dk-corpo

Les Echos: how DigitalKin is democratizing agentic AI for SMEs

Thu, 02 Jul 2026 15:34:53 GMT

The French business daily covers DigitalKin's mission to make rare, expensive expertise accessible to every company, especially SMEs, not just large corporations, through agentic AI.

Being able to afford rare, expensive expertise is still, today, a privilege of large corporations. Breaking that barrier is exactly why DigitalKin exists, and Les Echos just dedicated an article to it.

Why this coverage matters

Les Echos is France's leading business and economic newspaper, the reference read by executives, investors and decision-makers. Being selected by its editorial team is a meaningful signal: it reflects the growing recognition of agentic AI as a strategic shift for French companies, and of DigitalKin as one of its pioneers.

What the article covers

On June 30, 2026, the French business daily, Les Echos , published an article on our approach: custom agentic AI agents, our Kins , each built on the know-how of a real expert , to make high-level services accessible to SMEs that could never afford them before.

The article, written by Stéphane Frachet , Lyon correspondent, looks at how our Kins work , the experts and clients we work with, and the traction behind our model.

The conviction behind the story

For us, this coverage tells more than a product story. It tells a conviction, carried by cofounders Emmanuel Théry and Sébastien Deschaux since day one, and brought to life by a team of 11 building from Lyon for our clients: AI should multiply experts, not replace them .

That's what trusted AI means in the new AI economy : analysis reliable and traceable enough that a professional accepts to put their reputation behind it.

Read the full article on Les Echos (subscribers, french):
" DigitalKin démocratise les IA agentiques pour les PME"

About DigitalKin : Founded in Lyon in 2023, DigitalKin is a French deeptech startup specializing in agentic AI and causal AI . Its platform lets domain experts turn their know-how into autonomous AI services, accessible to every business. Learn more

DigitalKin Featured in the 2026 Auvergne-Rhône-Alpes French Startup Mapping

Tue, 30 Jun 2026 09:39:13 GMT

DigitalKin is a French deeptech startup pioneering causal and agentic AI from Lyon since 2023. The company is featured in the 2026 Auvergne-Rhône-Alpes French Startup Mapping , the annual reference cartography of the regional tech ecosystem co-published by France Digitale and Mesh Ventures , with the support of the local French Tech chapters: La French Tech Alpes and La French Tech Saint-Étienne Lyon .

What Is the 2026 Auvergne-Rhône-Alpes French Startup Mapping?

The Auvergne-Rhône-Alpes French Startup Mapping is a yearly cartography that identifies and references the most notable startups operating in the AURA region of France. It is co-published by France Digitale , Europe's largest startup and investor association, and Mesh Ventures , with the valuable support of the regional French Tech capitals: La French Tech Alpes and La French Tech Saint-Étienne Lyon . The 2026 edition highlights 1,311 regional startups that put innovation, technology and ambition at the heart of their business, showcasing the excellence of the Auvergne-Rhône-Alpes ecosystem. The mapping serves as a strategic reference for investors, institutions, large enterprises and ecosystem stakeholders across France and internationally.

Why DigitalKin Was Selected

DigitalKin, founded in Lyon in 2023 by Emmanuel Théry and Sébastien Deschaux, builds the French agentic AI platform where recognized professionals launch and commercialize autonomous AI services powered by their own domain expertise.

The company addresses high-stakes business missions - R&D literature reviews, regulatory compliance documentation, brand strategy, scientific analysis, innovation strategy, where generic AI tools fall short because the outcome requires professional judgment and experience, not just content generation.

The selection criteria for the mapping require startups to demonstrate a scalable model with breakthrough innovation or a strong technological differentiator, independent governance, and a headquarters in the AURA region.

Discover the full mapping and explore all the featured startups at francedigitale.org/publications/mapping-startups-aura-2026 .

What Makes DigitalKin Different: Causal AI and Expert-Endorsed Services

DigitalKin's core technology is a proprietary causal AI engine developed over six years of R&D. Unlike prompt-driven generative AI systems, Telos reasons backward from the objective to engineer the path toward the desired outcome, step by step. Every decision in the process is explainable. Every deliverable is auditable, sourced and traceable, reliable enough for domain experts to endorse it with their name and professional reputation.

Three technical pillars define the DigitalKin platform :

Telos (Causal AI Engine): Engineers the reasoning path from objective to deliverable.
The Kin (Personal AI Agent): Each user has a single AI agent, their Kin, that acts as a chief of staff: guardian of their data, strategist for their projects, executor of their daily missions, and gateway to the full ecosystem of expert services.
Model-agnostic, multi-infrastructure stack: DigitalKin's technology is not dependent on any single LLM provider. Designed for sovereignty, not dependency.

The DigitalKin Platform: How It Works

The DigitalKin Hub is the platform where two populations meet.

Expert professionals : consultants, researchers, senior specialists... productize their know-how into autonomous AI services that clients can access 24/7, at a fraction of traditional consulting costs, without any technical skills required. The expert endorses the service because DigitalKin's technology makes the output reliable enough to stake their reputation on.
Enterprises of any size access these expert-grade AI services for critical business missions without needing to prompt, code or manage an AI stack. Key use cases already in production include scientific literature reviews, R&D tax credit reports (Frascati-aligned, audit-ready from day one), continuous technology and market intelligence or brand positioning strategy.

DigitalKin in the French AI Ecosystem

DigitalKin is a member of the NVIDIA Inception program and has been backed by the CCI Lyon Métropole Saint-Étienne Roanne through Mesh Venture. The company was featured at VivaTech 2026 , where it publicly launched the Hub DigitalKin platform. The Auvergne-Rhône-Alpes region represents the second-largest concentration of AI startups in France after Île-de-France, accounting for 8% of the country's 1,114 AI startups according to the France Digitale AI Mapping 2026.

About DigitalKin Founded in Lyon in 2023 by Emmanuel Théry and Sébastien Deschaux, DigitalKin is a deeptech startup specializing in business-applied agentic AI, backed by CCI Lyon, Hub612 and member of the NVIDIA Inception program. Its mission: enabling experienced professionals to make their know-how accessible to all through autonomous, reliable, expert-grade AI services, without requiring any technical skills.

Expertise is no longer sold by the hour: DigitalKin reveals its platform at VivaTech 2026

Mon, 15 Jun 2026 21:43:03 GMT

The Lyon-based deeptech startup launches the Hub that lets professionals turn their know-how into an autonomous AI business, without a single developer.

At VivaTech 2026, Europe's largest innovation event, DigitalKin chose the Hub France IA Village to publicly unveil its platform for the first time: a complete technology allowing professionals to transform their intellectual know-how into an operational, commercial AI service, including client acquisition, expert-grade deliverable production and billing, without hiring a single developer or spending thousands on marketing.

The billable hour: expertise's glass ceiling

An expert, no matter how brilliant, can only serve one client at a time. Thirty years of know-how, capped by a calendar and a daily rate. The consequence is well known: high-level intellectual services only reach those who can afford them. An entire segment of the market that needs these services never gets access, for lack of budget, network or simply availability.

And the dominant AI focus on internal productivity changes nothing: gaining a few efficiency points does not break through a model where value remains indexed to time spent.

AI served coders first. Now it's the experts' turn.

AI makes it possible, for the first time, to separate know-how from the time of the person who holds it. But seizing this opportunity remains reserved for those who master the technology. Without technical skills, without a dedicated team, the barrier to entry is considerable for professionals who want to exploit AI's full potential in their field.

DigitalKin tackles both problems head-on. Invest in AI before it's too late, yes, but without a budget of hundreds of thousands of euros or a technical hire on one hand, and without compromising on quality, security and business results on the other.

"Today, a professional who wants to build an AI business has two options: invest hundreds of thousands of euros in a technical team, or spend months mastering AI at the expense of what they do best. We created the third way: the expert keeps developing their expertise, the technology does the rest," says Sébastien Deschaux, Chief Science Officer and cofounder of DigitalKin.

Two proprietary technologies, born in Lyon

DigitalKin opens two proprietary technologies to the market.

Telos, a causal reasoning AI engine built on 6 years of R&D, works the opposite of classic generative tools: instead of producing content and hoping it fits the need, Telos starts from the client's business objective to deliver high quality, not generic answers. The result: expert-grade, tailored deliverables, calibrated to a precise business challenge. Over 12,000 users, approximately 90% satisfaction, dozens of enterprise clients.
Agentic Mesh, a multi-agent, multi-model agentic AI architecture capable of running complex missions end to end. This technology is embodied by the Kins: autonomous AI agents that execute missions with full transparency, under the control of the user and the domain experts who transfer their know-how to them.

From scarce expertise to abundant expertise

The process is straightforward. A professional shares their domain expertise with a Kin that evaluates the commercial potential of that know-how, builds their AI service, generates their digital storefront, and handles production and billing for end clients. The entire cycle runs autonomously on the platform.

The target for each service creator: €30,000 in monthly revenue on the Hub, but above all, making their know-how available to every business that needs it, online, 24/7, at a price that changes the market.

What shifts for the market: high-level intellectual services, previously accessible only to an elite, become available to a vastly broader base of companies. Services already live or in development on the Hub include Scientific Advisory Board preparation for biotechs, CE marking clinical dossiers for medtechs, brand positioning for SMEs and startups, business model design for entrepreneurs, scientific literature reviews for R&D teams, and state-of-the-art analyses for R&D tax credit applications.

Applications open at VivaTech

DigitalKin is opening this capability first to a select circle of early partners: domain specialists, consultants, sector experts. Professionals can join the waitlist at VivaTech 2026.

Among the first partners, branding agency Nikita launches SpoK SME & Startups Branding: an AI-powered brand positioning service designed for SMEs. A strategic engagement that traditionally costs several thousand euros in consulting fees.

"At Nikita, we've positioned over 100 major brands in 30 years. We know what makes a brand strong. With DigitalKin, we're making that expertise accessible to every SME that needs it, not just those that can afford an agency. SpoK is our know-how, amplified by AI, for those who never had access before," says Karine Jamroszczyk, Brand Strategist at Nikita.

"Building an AI business is not about automating what you do internally and gaining a few productivity points. It's about amplifying the value you bring to your clients: making it more accessible, more available, more rigorous. Millions of professionals worldwide hold know-how that entire markets need. We give them the means to reach those markets, and in doing so, put their expertise at the service of everyone," says Emmanuel Théry, CEO and cofounder of DigitalKin

Meet DigitalKin at VivaTech Paris Expo Porte de Versailles | June 17-20, 2026 | Hub France IA Village | Hall 7 | Stand 2F14-005

DigitalKin officially listed in the Hub France IA 2026 cartography as a Trusted AI Partner

Sun, 22 Mar 2026 10:10:41 GMT

The official French AI startup and vendor mapping, presented at Bercy, now includes DigitalKin's agentic platform among the 876 companies shaping France's AI ecosystem.

How do you find the right AI partner for your business?

In a landscape where hundreds of solutions compete for attention, the question is harder than it sounds. The Hub France IA published its 2026 cartography of French AI startups and vendors, a reference mapping presented at Bercy on March 9 in partnership with the Direction Générale des Entreprises (DGE) , in front of nearly 300 ecosystem stakeholders. The initiative is part of the French government's "Osez l'IA" program , led by the Direction Générale des Entreprises (DGE) under the Ministère de l'Économie, des Finances et de la Souveraineté industrielle et numérique , designed to make the AI landscape more readable for businesses navigating their transformation.

DigitalKin: certified Trusted AI Partner

DigitalKin is proud to be officially listed in this sourcing and to receive the "Partenaire IA de Confiance" certification. Out of 876 innovative AI startups and vendors now structuring the French ecosystem, DigitalKin's agentic platform, the Hub, and its Kins are featured as a qualified solution for businesses looking for reliable, expert-grade AI.

What Makes DigitalKin Different

Kins are not generic AI assistants. They are autonomous AI agents built on real domain expertise. They handle complex strategic missions end to end: they reason, produce and deliver with full transparency, under the user's control. Each Kin is configured with a specific expert's methodology , vocabulary and quality standards, making it a true digital collaborator trained to the user's own requirements.

Services already live on the Hub include scientific advisory board preparation for biotechs, R&D literature reviews for researchers, brand positioning for SMEs, CIR tax credit applications, and business opportunity detection for expert entrepreneurs.

A necessary step for the French AI ecosystem

The Hub France IA cartography addresses a real market need: qualification. For companies evaluating AI partners, having an official, government-backed mapping of serious players reduces risk and accelerates decision-making. DigitalKin's inclusion confirms the maturity and reliability of its agentic technology in high-stakes professional environments.

The European dimension is coming next. Through Horizon Europe projects led by Hub France IA and partners including the DeployAI initiative , a European mapping is in preparation.

Try the platform

Discover the Kins already operational for clients at hub.digitalkin.ai, or learn more at digitalkin.com. The full 2026 cartography is available on the Hub France IA website .

DigitalKin was officially listed in the Hub France IA 2026 Cartography of AI Startups and Vendors, presented at Bercy, Paris, on March 9, 2026.

DigitalKin selected in the France Digitale 2026 mapping of French AI startups

Tue, 10 Feb 2026 09:15:53 GMT

DigitalKin is officially part of the France Digitale 2026 mapping of French AI startups . For any business searching for a trusted French agentic AI solution, this listing is a strong signal: DigitalKin is recognized as one of the companies structuring the French AI ecosystem.

What is the France Digitale AI startup mapping?

Every year, France Digitale , Europe's largest startup association with more than 2,000 members, publishes the reference mapping of French artificial intelligence startups . The 2026 edition, produced with the support of Sopra Steria Ventures , was unveiled at France Digitale's AI Day, just days before the AI Action Summit in India.

This is not an open directory. To be listed, companies must meet strict criteria: an independent governance, a scalable model built on genuine technological differentiation, and at least one commercial product based on AI technologies. The mapping is the qualified snapshot of who actually counts in French AI.

Why this selection matters

The 2026 edition confirms France as the leading AI ecosystem in Europe, with 1,114 identified AI startups, according to France Digitale. Together, these startups have raised close to €16 billion and account for nearly 50,000 jobs, making them a structural driver of the French economy. Being part of this select group places DigitalKin among the companies that investors, corporates and public institutions monitor when they look for serious French AI solutions.

For DigitalKin, this listing adds to a series of recent recognitions, including our Trusted AI Partner certification in the Hub France IA 2026 cartography , presented at Bercy with the Direction Générale des Entreprises.

Agentic AI: the trend the mapping confirms

One insight of the 2026 mapping stands out: the rise of agentic AI and highly specialized, business-specific AI agents is identified as the key dynamic of AI adoption today. After years where the main question was how to take AI in hand, turnkey and immediately operational solutions are now removing that barrier.

This is precisely the ground DigitalKin has been building on since 2023. Our Kins are autonomous AI agents, each built on the know-how of a real domain expert , executing complex professional missions end to end: scientific literature reviews , R&D tax credit applications, brand positioning, business analysis, scientific advisory board preparation.

A French answer to a global need

Businesses looking for AI solutions increasingly demand three things: reliability, traceability and sovereignty. As a French company, DigitalKin offers agentic AI that is white-box by design: every decision traceable, every result defensible, every deliverable endorsed by a real expert.

That's what we call trusted AI, and it's why organizations rely on our Kins for their high-stakes work.

Explore the interactive version of the 2026 mapping to see the full landscape of French AI startups.

Explore the mapping and our platform

The full announcement is available on the France Digitale website .

Discover the Kins already operational at hub.digitalkin.ai , or learn how to create your own AI service built on your expertise, for the benefit of your team and your clients.

About DigitalKin
Founded in Lyon in 2023, DigitalKin is a French deeptech startup specializing in agentic AI. Its platform lets domain experts turn their know-how into autonomous AI services, accessible to every business. Its mission: enabling experienced professionals to make their know-how accessible to all through autonomous, reliable, expert-grade AI services, without requiring any technical skills.Backed by CCI Lyon, Hub612 and member of the NVIDIA Inception program. Learn more

In 2026, your expertise deserves better

Thu, 01 Jan 2026 21:53:18 GMT

What happens when everyone uses the same AI? Discover how the Kins protect your singularity in 2026.

What happens when everyone uses the same AI?

This question is the driving force behind DigitalKin . For two years, we have been building an ecosystem where human expertise and technological power reinforce one another, moving away from standardisation. We created the Kins : expert AI agents designed to encapsulate and amplify your unique skills for the rest of the world.

In 2026, we are going further:

The DigitalKin Hub is becoming the benchmark for experts who want to differentiate themselves durably. ‍
Expansion of the ecosystem toward complex and highly strategic business processes. ‍
Reinforced support for companies in their quest for singularity.

The only remaining challenge? Spreading the Kins widely around us, and you.
Thank you for being part of this adventure

How do our AI agents for R&D and Innovation transform your daily work?

Our expert AI agents are not just standard tools; they are designed for scientific excellence. Already adopted by over 1,500 users in 2025, they are evolving to meet the most stringent requirements of research and engineering.

For your R&D projects, our Kins now feature advanced capabilities:

Literature Review and SOTA (State of the Art): Exhaustive and systematic analysis of scientific databases, patents, and academic publications. ‍
Academic Traceability: Every statement is referenced with rigorous precision, making the work fully auditable. ‍
R&D Tax Credit Support: A valuable partner for establishing the state of the art required for your R&D tax credit claims (such as CIR in France). ‍
Real-time Collaboration: Your Kin’s report adjusts instantly based on your feedback.

The result? Researchers and engineers are freed from weeks of manual literature search to focus on interpretation and innovation.

They talk about us: A committed ecosystem

In 2025, DigitalKin was supported and showcased by major players in the innovation landscape.

Toward an exceptional 2026

The entire DigitalKin team wishes you a brilliant year alongside your future digital colleagues. It is time to release the full potential of your expertise and create the impact you deserve.

DigitalKin's CEO on Lyon Décideurs: "Making expertise accessible and abundant"

Wed, 31 Dec 2025 21:29:47 GMT

Emmanuel Théry shares DigitalKin's vision on Le Journal Éco - Lyon Décideurs - building a digital workforce that encodes real expert know-how into autonomous AI agents.

Emmanuel Théry, CEO and cofounder of DigitalKin, was the guest of Jean-Pierre Vacher on Lyon Décideurs' Journal Éco to discuss how the Lyon-based startup is redefining AI for businesses.

Beyond AI assistants: a true digital workforce

"At DigitalKin, we develop AI agents capable of executing complex tasks in full autonomy, with unmatched reliability," Théry explained. The distinction matters: where others build AI assistants, DigitalKin builds a genuine digital workforce. "Our agents already analyze hundreds of scientific publications simultaneously for major pharmaceutical companies, with expert-level precision."

Expertise made accessible and abundant

Théry laid out the company's core ambition: "Our goal is to make expertise accessible and abundant. By encoding the most specialized human know-how into autonomous agents, we enable companies to multiply their expertise without limits."

This vision goes beyond productivity gains. It is about fundamentally changing who can access world-class expertise, and at what scale. A scientific literature review that once took months can now be delivered in minutes. The quality standard is maintained by the expert whose methodology is encoded at the core of each Kin.

A startup that moves fast

Founded in 2023 in Lyon with a team of ten, DigitalKin has already attracted major clients across pharmaceuticals, industrial R&D and scientific research. A member of the NVIDIA Inception program and winner of French Tech Seed, the company is preparing to scale commercially with industrial, scientific and medical organizations.

Emmanuel Théry spoke on Le Journal Éco, Lyon Décideurs. Full video interview available on lyondecideurs.com. Originally published December 31, 2025.

AI and ethics: inside Lyon's rising agentic AI ecosystem

Thu, 18 Dec 2025 21:16:53 GMT

Discover the interview >

The Lyon deeptech turning AI into expert AI agents that are reliable, personalized and truly useful

Tue, 25 Nov 2025 21:25:57 GMT

Lyon Entreprises met DigitalKin's CEO to understand how the startup is building a new generation of AI agents that amplify human expertise instead of replacing it.

In Lyon's fast-moving tech landscape, DigitalKin is one of those companies that moves fast. Founded in 2023, this ten-person deeptech startup carries an ambitious vision: let anyone transform their expertise, whether scientific, technical or industry-specific, into a tailor-made AI agent. A deceptively simple idea that addresses one of the biggest challenges in AI today: making it truly reflect the singularity of the people and organizations that use it.

Why most AI projects fail

For DigitalKin's team, the diagnosis is clear. Most AI projects fail not because of a lack of technology, but because of a lack of personalization. The mistake, they say, is forgetting context, specificity, the unique know-how that defines every organization. The companies that succeed are those that start from their DNA, from what makes them different, and amplify it through artificial intelligence. The team's core message: AI should not flatten. It should reveal.

A platform already in action

It is with this philosophy that DigitalKin recently launched its new platform. From digitalkin.com, users can already test a first agent capable of producing, in full autonomy, scientific literature reviews. This is work normally reserved for experts: analyzing thousands of studies, verifying sources, cross-referencing data, understanding what science actually says about a precise subject, such as the efficacy of a cancer drug. The agent does it step by step, without hallucinations, drawing on published literature, patents and scientific publications. A rare promise in a world where approximate answers are still far too common.

Already adopted by major organizations

Though still in its early days, DigitalKin already counts several major clients. In the Lyon region, the company works with Groupe SEB, Boiron laboratories and several pharmaceutical companies. Healthcare is becoming a key growth sector, driven by use cases where precision and reliability are non-negotiable. For the team, the coming months will be decisive: the goal is clear, accelerate commercially with industrial, scientific and medical organizations.

AI that amplifies, not replaces

DigitalKin is positioning itself as a leader in expert AI: reliable, contextualized, and above all useful. AI that does not replace human knowledge, but amplifies it at a scale previously impossible.

Originally published by Lyon Entreprises, November 21, 2025. Full video interview available on lyon-entreprises.com.

What's Next? Towards an Open Standardization of Distributed Intelligence

Thu, 20 Nov 2025 09:00:00 GMT

Launched open source by Anthropic in November 2024, the Model Context Protocol (MCP) represents more than just technical progress: it is a strategic turning point in the architecture of artificial intelligence. In less than a year, the protocol experienced rapid adoption — over 1,000 MCP servers developed by February 2025 , and an ecosystem exceeding 5,500 servers by autumn.

This spectacular growth reflects a structural shift: the transition from a human-user-centric tool-AI to an agent-AI embedded in distributed ecosystems .

Until now, language models operated in siloed environments. With MCP, these models acquire an interoperation capability : they can act in the digital world, interface with external systems, orchestrate data flows, and collaborate with other agents.

This evolution paves the way for an agentic web , where interactions are no longer limited to human-machine interfaces but are deployed between autonomous agents capable of reasoning, learning, and cooperating. The protocol becomes the common language of this new infrastructure layer, connecting AIs to each other as TCP/IP once connected computers.

The MCP inaugurates a cognitive interconnection infrastructure : each server exposes capabilities, each agent discovers and uses them, and each interaction is traceable, verifiable, and governed.

In this perspective, MCP plays the same role for the agentic era as HTTP for the web or Kubernetes for the cloud : an invisible yet structuring standard that enables large-scale coordination.

Anticipating the future of the MCP requires recognizing its systemic scope. The protocol does not just transform how AIs are integrated; it reconfigures economic models, technological power balances, ethical and regulatory frameworks, and digital sovereignties.

Economic models , by fostering the creation of interoperable ecosystems and agent marketplaces.
Technological power balances , by redefining dependencies between cloud, data, and AI players.
Ethical and regulatory frameworks , by raising the question of governance for decisions made by autonomous agents.
Digital sovereignties , because an open protocol controlled by a few actors can become a major geopolitical lever.

The MCP is part of the trajectory of major digital coordination infrastructures: those that, by establishing themselves as de facto standards, shape the global economy for decades . Its rise announces a future where AIs will no longer be isolated tools but intelligent partners embedded in fluid, governable, and potentially planetary networks.

Protocol Evolution Perspectives: Modularity and Cross-Vendor Interoperability

Extensions, Plug-ins, and Standard Evolution

The Model Context Protocol (MCP) is currently based on a client-server architecture based on JSON-RPC 2.0 , directly inspired by the Language Server Protocol (LSP) .

This modular design is a deliberate choice: it guarantees the protocol native extensibility , meaning the ability to evolve without breaking compatibility.

Each new functional building block can be integrated as an extension, without challenging the core of the protocol—an essential approach for a standard meant to last.

‍

During the September 2025 MCP summit , Anthropic and its partners presented an ambitious roadmap for 2025–2026, articulated around four priority evolution areas.

1. Multimodality and Streaming: Towards Perceptive Agents

One of the most anticipated projects concerns the native support for multimodality .

‍

The objective: to allow agents to process text, audio, video, and visual data simultaneously, while maintaining contextual coherence between these streams.

‍

This evolution relies on two key innovations:

Native bidirectional streaming , for continuous and reactive exchanges between agents and servers.
Dynamic chunking , enabling the segmentation of large data volumes (e.g., videos, time series, complex logs) to facilitate incremental processing.

‍

With these capabilities, the MCP will become a true cognitive channel for agents capable of understanding a meeting, analyzing visual signals, or synchronizing events in real time—a decisive step towards AI embedded in operational environments.

2. MCP Registry: The Backbone of Trust

Launched in preview version in September 2025, the MCP Registry aims to become the single source of truth for the discovery, verification, and distribution of MCP servers.

This centralized registry brings several major innovations:

Automated discovery : each agent will be able to identify compatible servers without manual configuration.
Certification and signing : a trust verification mechanism authenticates servers via public keys.
Meta-indexing : the registry stores deployment metadata (version, dependencies, maintainers, compatibility).

The goal is to provide companies with a secure, traceable, and governable ecosystem , avoiding the proliferation of unverified servers—an essential step for the industrial maturity of the protocol.

3. Enhanced Authentication and Authorization

One of the main challenges of the MCP remains the security of access flows between agents, clients, and servers.

The roadmap includes the full integration of modern security standards:

OAuth 2.1 for secure authorization,
Dynamic Client Registration (DCR) for automated registration of client applications,
Enterprise SSO (Single Sign-On) via integrations with Okta , Azure AD , or Auth0 ,
And the explicit separation between resource servers and authorization servers , to reduce the risks of privilege escalation.

‍

Although these developments are already underway in the specification published in June 2025, challenges persist around compliance with the security policies of large enterprises —particularly for decentralized identity management and fine-grained permission delegation (hybrid RBAC/ABAC).

4. Reference Implementations and Multi-Language Interoperability

The success of the MCP relies on its ecosystem neutrality . Following the Python and TypeScript SDKs, implementations in Java, Go, Rust, and C# are now in development. Each is accompanied by automated compliance test suites , guaranteeing that servers and clients strictly adhere to the specification.

‍

These tests, published as an "MCP Compliance Kit," will allow companies to validate their internal implementations before deployment—a prerequisite to prevent standard fragmentation.

Open Governance and Improvement Process

To ensure its sustainable evolution, the MCP relies on an ** open community governance model**. The core of this model is the Specification Enhancement Proposal (SEP) —a process inspired by PEPs (Python Enhancement Proposals) and IETF RFCs .

‍

A proposal's life cycle follows six transparent steps:

Proposal → Draft → Provisional → Prototyping → In Review → Accepted.

‍

Each SEP must be sponsored by a core maintainer and validated collectively after public discussion. Working Groups (security, multimodality, compliance, transport) and Interest Groups (user companies, academia, open source) meet regularly to guide decisions.

‍

Bi-weekly meetings of core maintainers and the systematic publication of minutes ensure a high level of transparency.

This distributed model has a dual objective:

Avoid the capture of the protocol by a dominant actor,
And guarantee technical coherence in a rapidly expanding ecosystem.

MCP as a Building Block for an Agentic Web: Cooperative AIs and Interoperable Mesh

The Emergence of the Agentic Mesh

The concept of the Agentic Mesh extends the Model Context Protocol (MCP) towards a distributed architecture where several specialized agents collaborate within a coordinated network. The idea is inspired by service meshes (Istio, Linkerd) in the cloud world but transposes their supervision and orchestration logic to autonomous cognitive systems .

‍

The Agentic Mesh is not just a technical evolution of the MCP: it embodies its maturity, where AI no longer just responds, but orchestrates, coordinates, and learns collectively .

Composability and Modularity
Each agent, model, or tool becomes an interchangeable node of the network.
It can be added, replaced, or updated without modifying the other components , thanks to interface standardization via MCP.
This modularity offers a dual advantage: agility of evolution (incremental addition of capabilities) and scalability (natural horizontal scalability).
Distributed Parallel Reasoning
Where early AI architectures centralized thought in a single monolithic model, the mesh distributes tasks among specialized agents working in parallel.
An agent can, for example, extract data while another synthesizes it, and a third checks for consistency.
This asynchronous operation improves performance and allows for domain specialization : finance, health, logistics, scientific research...
Logical Decoupling and Layered Governance
The mesh separates responsibilities between logic, memory, orchestration, and interface .
Each agent operates autonomously while sharing a synchronized common context .
The whole is documented in a behavior logging system , which journals every tool invocation, every error, and every decision—an essential approach for traceability and compliance.
Vendor Neutrality
True to the open standard spirit, the Agentic Mesh favors open protocols (MCP, Google's A2A) over proprietary APIs .
Components can thus be replaced without global reconfiguration, ensuring interoperability and sovereignty .

‍

Concrete implementations like AgentMesh or multi-server MCP architectures already demonstrate the viability of this model.

Interoperability and Competing Protocols

The MCP is currently the most advanced standard for agent-to-tool communication, but it is not alone in the field. Several competing or complementary initiatives seek to cover other layers of the agentic ecosystem.

Announced in April 2025 with the support of over 50 partners (Salesforce, MongoDB, PayPal, etc.), the A2A protocol focuses on inter-agent communication rather than tool access. It also uses JSON-RPC 2.0 over HTTPS and introduces an agent card mechanism enabling the discovery of capabilities in a decentralized manner.

‍

Google presents A2A as complementary to MCP :

MCP manages the connection between agents and resources (data, APIs, tools).
A2A manages the coordination between autonomous agents.

‍

In practice, the two overlap in certain use cases, creating a healthy tension between convergence and specialization.

Web of Agents: A Federated Architecture

The Web of Agents project (arXiv:2505.21550) formalizes this vision of an interoperable agentic ecosystem at the web scale . Its minimal architecture rests on four pillars:

Agent-to-Agent messaging —standards for communication between autonomous entities.
Interaction interoperability —standardized data formats and exchange protocols.
Distributed state management —synchronization mechanisms to maintain coherence between agents.
Agent discovery —open registries allowing the identification and qualification of available capabilities.

‍

This work joins the initiatives of the W3C "Autonomous Agents on the Web" Community Group , which explores how to adapt web standards (HTTP, WebSockets, RDF, DID) to the needs of distributed intelligence.

‍

The common objective: to avoid fragmentation into incompatible silos and lay the foundation for an "Internet of Agents" —an open, secure, auditable, and evolutive network, where humans and AI cooperate on an equal protocol footing.

Regulatory, Ethical, and Technological Sovereignty Challenges

Distributed Governance and Responsibility

The increasing autonomy of AI agents, made possible by the Model Context Protocol (MCP) , disrupts traditional frameworks of responsibility and governance.

As soon as several agents collaborate within the same system—for example, a financial agent , an HR agent , and a logistics agent coordinating an operation—a central question arises: who is responsible in case of error, prejudice, or non-compliant decision?

‍

The MCP, as a communication infrastructure, does not intrinsically define responsibility . It provides the channel, not the ethical or legal framework. This structural gap places governance at the heart of the agentic era's challenges.

Legal Responsibility and Regulatory Compliance

New regulations, such as the EU AI Act (effective August 2024), impose strict requirements on high-risk AI systems regarding transparency, traceability, and risk management .

‍

To be compliant, an agentic architecture based on MCP must:

Trace all inter-agent interactions thanks to detailed audit logging,
Allow access revocation to a compromised agent or server,
And make decisions explainable within a distributed reasoning process.

‍

Current MCP specifications only partially integrate these functions.

The Compliance Working Groups are now working on a major project: adding compliance by design modules, allowing the direct integration of audit, access, and traceability policies into the protocol itself.

‍

The logic of “governance by design” is thus becoming essential. Access, consent, and privacy policies must be encoded in the MCP configurations , not added a posteriori.

‍

New approaches are emerging, such as Explainable MCP , which aims to record not only the executed actions but also the agents' underlying reasoning —an equivalent of the "intention log" within the agentic network.

Bias, Fairness, and Transparency

Like any technology based on learning models, systems based on MCP can amplify biases present in training data or in the tools they use.

An HR agent connected to a biased recruitment system could reproduce—or even extend—large-scale discrimination.

‍

To address this, MCP actors are exploring several levers:

Systematic bias audits , carried out at regular intervals on servers and agents,
Integrated Fairness checks in MCP workflows, validating equity criteria before execution,
Transparent decision chains , where each reasoning step is logged with its contextual metadata (source, date, justification).

‍

The MCP Registry could play a structuring role by certifying ethical servers —i.e., those compliant with non-discrimination, privacy, and regulatory compliance criteria. Eventually, a “Fair MCP” label could be imagined, attesting to the ethical compliance of a server, similar to ISO or SOC certifications in the cloud.

Technological Sovereignty and Strategic Independence

Beyond ethical issues, the challenge is also geopolitical.

Europe has made digital sovereignty the core of its AI strategy, notably through two initiatives:

the AI Continent Action Plan (April 2025),
and the Apply AI Strategy (October 2025).

‍

These plans aim to reduce dependence on American (OpenAI, Microsoft, Google) and Chinese ecosystems by developing an autonomous AI value chain : data, computing, models, and governance.

‍

The MCP, although born at Anthropic (an American startup), paradoxically offers a unique opportunity for Europe . As an open standard , it is not captive to a single actor and can serve as the foundation for a sovereign agentic ecosystem .

‍

Concretely, Europe could:

develop public MCP servers for health, education, administration,
build certified European registries , guaranteeing ethical and regulatory compliance,
and encourage auditable open source implementations , hosted on sovereign infrastructures.

‍

The “AI First” strategy promoted by the European Commission already encourages companies to integrate AI as a lever for competitiveness and productivity.

The MCP, by facilitating the integration of AI agents into existing business systems, could become the natural accelerator of this transformation —provided that European actors master the entire technological chain : models, computing infrastructures, and governance protocols.

Role of Open Source and Industrial Alliances

Community Dynamics and Industrial Adoption

The Model Context Protocol (MCP) did not just benefit from good technical design; it primarily found its strength in a strategy of radical openness . From its launch, Anthropic made the decisive choice to make the protocol open source , allowing any actor—individual developer, startup, or large group—to contribute without prior authorization . This decision triggered a distributed innovation dynamic that quickly transformed into a genuine ecosystem movement .

‍

In less than a year, over 5,500 MCP servers have been developed, a figure that illustrates the speed of standard diffusion and the maturity of its technical community . This vitality is reinforced by the progressive convergence of several major industrial alliances.

Predictions and Future Trajectories (2025–2030)

Standardization and Maturity (2025–2026)

The next two years will mark the stabilization of the standard .

According to the community roadmap, version 2.0 of the MCP , expected for November 25, 2025 , will integrate three structuring evolutions:

native multimodality (text, audio, video, signals),
bidirectional streaming for continuous data flows,
and enhanced authentication compliant with OAuth 2.1 and DCR.

‍

This stable version will constitute the necessary technical basis for large-scale production adoption .

‍

In parallel, a set of compliance frameworks will emerge:

standardized test suites to validate implementations,
MCP server certifications guaranteeing security and interoperability,
and a centralized registry aggregating server metadata, facilitating component discovery and verification.

‍

‍Mirroring the role played by the CNCF Landscape in cloud native, the MCP Registry will become the central hub of the agentic ecosystem .

Proliferation of Agents and Multi-Agent Orchestration (2026–2028)

Starting in 2026, the generalization of Agentic AI will transform the very nature of information systems. Multi-agent architectures orchestrated via MCP will become the norm in complex environments.‍

‍

New organizational models will emerge:

a coordinator agent ("Chief of Staff") managing planning,
surrounded by specialist agents (Coder, DevOps, Data Analyst, Legal Advisor, etc.),
all interconnected via MCP and exchanging in real-time within a distributed mesh.

‍

‍These Agentic Meshes will become the reference architectures for distributed cognitive systems, with fluid, traceable, and governable communication between each agent.‍

‍

Major cloud players will quickly follow suit:

AWS Lambda , Google Cloud Run , and Azure Functions will natively integrate the MCP protocol,
enabling the serverless deployment of agents capable of mutual orchestration without going through proprietary APIs.

‍

This convergence between cloud and agentic mesh will mark the fusion between infrastructure computing and adaptive intelligence .

Consolidation and Mature Ecosystem (2028–2030)

Between 2028 and 2030, the MCP will reach its industrial maturity .

The ecosystem will enter a phase of rationalization comparable to that experienced by the cloud in the early 2020s:

The most efficient orchestration patterns will be standardized.
Reusable libraries of servers and agents will emerge.
Multi-tool servers , capable of exposing several capabilities in a single module, will replace single-function implementations.
Companies will turn to packaged and certified solutions , prioritizing security and compliance over artisanal flexibility.

‍

‍The MCP will also evolve through integration with other structuring technologies:

Blockchain and smart contracts for traceability and trust between agents.
Edge computing to bring agents closer to the data they process (industry, health, IoT).
Post-quantum cryptography to secure long-term communications.

‍

These synergies will transform the MCP into the backbone of a distributed intelligence ecosystem , interconnecting billions of agents and heterogeneous systems.

Economic Value and Industrial Impact

The economic impact of the MCP will be considerable. According to McKinsey , interoperability between AI systems—of which the MCP is the keystone—could generate up to 300 billion dollars in annual added value in critical infrastructures by 2030.‍

‍

This value will primarily come from:

reduced integration costs,
automation of high-cognitive-intensity tasks,
and mutualization of models across partner ecosystems.

‍

‍But maturity will not come without risks.

According to Gartner , by 2027, more than 40% of agentic AI projects could be abandoned due to a lack of clear return on investment or adequate governance .

‍

The MCP, by reducing integration complexity and structuring interaction flows, has the potential to reverse this trend— provided that companies adopt a disciplined approach :

define tangible KPIs (time saved, resolution rate, value generated),
measure traceability and operational performance of agents,
and invest in the training and human supervision of systems.

Conclusion: Towards an Open and Governed Agentic Web

The Model Context Protocol (MCP) is more than just a technical standard: it is a civilizational infrastructure for distributed intelligence—the equivalent, for the agentic era, of what HTTP was for the content web.

It constitutes the interoperability layer without which artificial intelligence systems would remain confined to isolated environments, unable to cooperate or self-organize.‍

‍

Its evolution towards a mature ecosystem—integrating multimodality , trust registries , distributed governance , and cross-vendor interoperability —will determine the very form of the global agentic web .

Its structuring will determine the possibility of an intelligent agent network capable of collaborating, learning, and reasoning on a planetary scale without compromising security or human responsibility.‍

‍

The stakes of the MCP far exceed the field of software engineering. The goal is now to build an open, secure, and ethical agentic web , where artificial intelligences can interact while respecting:

data sovereignty ,
protection of private life ,
and the democratic values on which our societies are based.

‍

In this new paradigm, transparency, traceability, and shared governance are no longer options but foundations. The question is no longer just how agents will communicate, but under whose authority, according to which rules, and serving what ends.‍

‍

The MCP is not an end, but the foundation of a profound recomposition of digital systems. It opens the way to a new era—one where AI agents cooperate in governed networks , capable of collective reasoning, distributed action, and iterative learning.‍

‍

Organizations that master this infrastructure—while anticipating its ethical, regulatory, and geopolitical challenges —will occupy a determining position in the decade ahead. They will not merely use AI: they will become its systemic architects .

The Founding Principles of the Agentic Mesh

1. Google Agent-to-Agent (A2A)

2. Other Emerging Protocols

Agentica (WrtnLabs) : aims to reduce integration costs by 50% by relying on lightweight models and simplified servers.
Universal Tool Calling Protocol (UTCP) : removes the proxy architecture to reduce latency by half , at the cost of losing flexibility.

‍

These initiatives reflect a classic dialectic in the history of standards : MCP favors universal interoperability and robustness, while its alternatives explore raw performance or implementation simplicity .

3. Towards a Multi-Protocol Consensus

In the long term, the ecosystem could stabilize around a multi-layer architecture :

MCP for data and tool access,
A2A for inter-agent coordination,
And specialized protocols (UTCP, Agentica) for vertical or embedded uses.

‍

This convergence would recall the evolution of the Internet, where HTTP, FTP, and SMTP coexisted while fulfilling distinct functions.

‍

The adoption by Microsoft of native MCP support in Windows 11 , and its integration into GitHub Copilot and Copilot Studio , suggests that a common foundation is emerging, upon which sectoral extensions will be grafted.

Microsoft: The Copilot–MCP Convergence

Microsoft is today one of the pillars of MCP adoption. The protocol is now natively integrated into Windows 11 , GitHub Copilot , and Copilot Studio , creating a continuum between development environments and conversational agents.

The Dataverse service, Microsoft's enterprise database, exposes an official MCP server , allowing agents to directly access business datasets without going through proprietary APIs.

Hugging Face: Open Standardization of AI Knowledge

Hugging Face has published its own official MCP server , providing access to the models, datasets, Spaces, and articles hosted on its platform.

This integration, compatible with Claude Desktop , VS Code , and Cursor , transforms the Hugging Face Hub into a universal contextual reservoir for MCP agents—a space where AIs can not only draw data but also interact with learning artifacts.

Google: Prudent but Strategic Adoption

Google has chosen a gradual approach.

While Gemini and Google Workspace already implement MCP-compatible protocols, the company continues to promote its own standard, A2A (Agent-to-Agent) , in parallel. This cohabitation illustrates a strategy of balance: supporting interoperability while retaining a lever of influence on the inter-agent communication layer.

OpenAI: The Interoperability Turn

In 2025, OpenAI integrated MCP support into its Agents SDK , allowing GPT-4 and its successors to directly invoke tools via MCP servers. This move marks a profound strategic change: the market leader (with 35% enterprise adoption) now recognizes the value of a cross-vendor protocol —an implicit recognition that market growth will come through standardized coopetition , not proprietary lock-in.

Developer Ecosystem: The Living Base of the Standard

The most popular development tools— Cursor, Cline, Zed, Replit, Sourcegraph —have integrated the protocol. Around them, an ecosystem of community registries (mcp.so, MCP Market , PulseMCP ) has been established, facilitating the discovery, rating, and sharing of servers.

These platforms function as the "App Stores" of the agentic world, where trust, traceability, and reputation become the new drivers of adoption.

Linux Foundation and Multi-Stakeholder Governance

To date, no official announcement confirms the hosting of the MCP under the aegis of the Linux Foundation or an equivalent structure.

However, many observers anticipate that as the protocol consolidates, a multi-company consortium will emerge to guarantee its neutrality, governance, and sustainability .

‍

This future consortium could play a role analogous to that of:

the Cloud Native Computing Foundation (CNCF) for Kubernetes,
or the OpenSSF for open source software security.

Its missions would be threefold:

Coordinate development efforts for reference implementations,
Finance the maintenance and security of the protocol,
Arbitrate technical evolutions and conflicts between contributors.

The current governance model— centered on Anthropic but open to external contributions—will have to evolve towards a collegiate structure .

This transformation will be crucial to credential the MCP as a global trust infrastructure , and not as a standard dominated by a private actor.

Security, Supply Chain, and Trust Registries

The open and distributed nature of the MCP, while promoting innovation, also creates a new field of vulnerabilities . In theory, anyone can publish an MCP server. In practice, this opens the door to supply chain risks : compromised, malicious, or simply misconfigured servers.

‍

A recent study highlighted the extent of the problem:

43% of tested implementations exhibited command injection vulnerabilities,
30% were exposed to SSRF flaws,
22% allowed arbitrary file system access .

‍

To respond to these threats, the developing MCP Registry will have to integrate:

certification mechanisms based on digital signature,
automated security audit ,
and server reputation systems (trust scoring).

‍

Open source tools like mcp-scan are already emerging to detect vulnerabilities, dangerous configurations, or tool poisoning attempts (malicious injections hidden in tool metadata).

‍

The open source ecosystem plays an essential role as a collective watch here. Initiatives like GitHub Security Advisories , OWASP AI Security Project , or AI Supply Chain SIG contribute to strengthening security practices around the protocol.

‍

However, companies will also have to adopt rigorous internal procedures :

whitelisting of authorized MCP servers,
automated security scans in CI/CD pipelines,
non-production sandbox testing before any deployment.

‍

This combination of community tooling and corporate governance will shape the foundations of a secure, certifiable, and sustainable MCP ecosystem .

The Model Context Protocol (MCP) is entering a decisive structuring phase.

After the euphoria of early adoption, the 2025–2030 decade will see the emergence of an era of consolidation and industrialization , where the protocol will become a true global cognitive infrastructure .

How MCP Transforms AI Architectures?

Mon, 20 Oct 2025 09:00:00 GMT

The Model Context Protocol (MCP) , launched by Anthropic in November 2024 , is rapidly establishing itself as the universal standard for connecting artificial intelligence systems to the real digital world. Designed as a true "USB-C for AI" , this open and extensible protocol is revolutionizing the way Large Language Models (LLMs) interact with data, tools, and enterprise environments.

In just a few months, MCP has garnered massive adoption: OpenAI , Microsoft , Google DeepMind , Cloudflare , and MongoDB have integrated it into their platforms. The ecosystem now exceeds 5,500 active servers , and the twenty most popular already generate over 180,000 monthly searches — proof of global enthusiasm for this new infrastructure.

From Monolithic SaaS to Agentic Mesh: A New Workflow Architecture

The Old Paradigm: Fragmented Integrations and Monoliths

Before MCP, connecting an AI application to external data sources or tools was a technical headache. Each connection required specific development, leading to an N×M integration problem : to connect N AI systems to M external services, N×M different connectors had to be built.

This logic resulted in:

Exponential maintenance costs ,
A proliferation of data silos ,
And chronic technical debt , with each API having its own syntax, constraints, and update cycle.

Monolithic SaaS applications exacerbated this rigidity: their static APIs required manual implementation of each endpoint. Result: AI agents remained context-blind and unable to adapt dynamically — the exact opposite of the cognitive flexibility promised by agentics.

The New Paradigm: Agentic Mesh Orchestrated by MCP

MCP introduces a new workflow architecture , based on a distributed network of interoperable agents, tools, and servers . Instead of a multitude of isolated integrations, we are witnessing the birth of a coherent, self-discoverable mesh , relying on three key innovations:

Dynamic Capability Discovery — MCP servers expose their functionalities (tools, resources, prompts) via a standardized protocol. An AI agent can automatically query a server to discover what it can do — without manual documentation. This is the equivalent, for artificial intelligence, of hardware "plug-and-play" : a new tool plugged in becomes immediately usable.
Bidirectional and Stateful Communication — Unlike REST APIs, based on stateless requests, MCP maintains a persistent session between the client and the server. Agents can thus conduct continuous dialogues with an external system: querying a database, analyzing results, refining the request, all while preserving context. This makes multi-turn conversations and progressive contextual learning possible.
Multi-Agent Orchestration — MCP provides the communication layer necessary for the cooperation between specialized agents . A search agent can invoke a code analysis agent, which itself solicits a documentation agent. The protocol ensures the synchronization, security, and coherence of these interactions, forming a true distributed intelligence .

Transformed Workflow: From Human to APIs via AI

The introduction of MCP disrupts the traditional pattern of digital interactions:

Before: The Classic Flow

Human → Application → Hardcoded API → Database → Response

With MCP: A Contextual and Intelligent Flow

Human → AI Agent (MCP Client) → MCP Server → API / Tool → Database → Enriched Context → AI Agent → Human

This model radically transforms the nature of the dialogue between the user and information systems. The agent becomes an intelligent orchestrator , capable of executing entire workflows across multiple systems.

For example, when a user asks: "What is my savings capacity this month?"

The AI agent, via MCP, can:

Query multiple banks (via DSP2 APIs),
Aggregate and normalize the data,
Detect anomalies or trends,
And provide a clear and contextualized summary.

All of this without the user manually managing OAuth tokens , API authorizations, or integration logic.

Concrete Benefits: Modularity, Auditability, Traceability, Scalability

The Model Context Protocol (MCP) does not just standardize exchanges between AI agents and external systems: it transforms software architectures by bringing four major benefits — modularity , auditability , traceability , and scalability .

These principles, derived from the best practices of software engineering, are finally becoming applicable to the world of artificial intelligence.

Modularity: Microservices Architecture for AI

MCP introduces a modular approach where each MCP server encapsulates a specific capability:

Access to Slack or Microsoft Teams ,
Querying PostgreSQL or Snowflake ,
Automated web navigation via Playwright ,
Or interaction with internal tools.

Each server thus becomes an independent functional block , reusable and composable. This architecture brings several decisive advantages:

Reusability : an MCP server developed for one project can be shared between teams or departments without rewriting.
Simplified Maintenance : an update or security patch is instantly applied to all connected clients.
Isolation of Responsibilities : each server can be developed, tested, and deployed independently, following the proven principles of microservices .

Concrete Use Case : At Block (formerly Square) , this modular approach allowed for the development of internal servers for Snowflake , Jira , Slack , and Google Drive , all accessible via a unified agent named Goose .

Result: a 50 to 75% reduction in time spent on common engineering tasks and a drastic acceleration of operational productivity .

Auditability and Traceability: End-to-End Visibility

The MCP architecture natively integrates observability mechanisms essential for businesses. Each interaction between an agent and a tool can be traced, logged, and analyzed with precision.

The three key components of this observability are:

Structured Logs : each tool invocation records its input parameters, output, execution duration, and associated metadata.
Distributed Tracing : thanks to integration with OpenTelemetry , it becomes possible to follow a request through multiple agents, MCP servers, and external APIs. This allows visualization of the complete execution chain — from the initial prompt to the final response — and instant identification of bottlenecks or errors.
Enhanced Auditability : in regulated sectors (finance, health, defense), MCP allows for the reconstruction of the complete history of decisions : Which agent accessed which data, when, and under what permissions?

This level of transparency, unthinkable in traditional AI architectures, becomes a key requirement for companies wishing to industrialize AI in a compliant and responsible manner .

Observed Production Results : Deployments based on MCP have recorded an average 60% reduction in incident detection time and 75% improvement in error recovery rate , thanks to this native and distributed observability.

Scalability: From Local Experimentation to Cloud Deployment

MCP was designed to support scaling, from individual experimentation to massive deployment. It supports three complementary deployment models :

Workstation (local STDIO) — The server runs locally on the developer's machine. Ideal for prototyping, quick testing, or tools requiring local access (files, IDE).
Managed (containerized) — MCP servers are deployed in orchestrated containers (like Kubernetes ), ensuring isolation, horizontal scalability, and high availability. This is the recommended mode for production environments.
Remote (HTTP + SSE or Streamable HTTP) — Servers expose their capabilities via HTTP, allowing multiple distant clients to connect to them. This model favors geographic distribution and integration into multi-cloud infrastructures.

Since May 2025 , "remote" deployments have seen a growth of over 400% , signaling large-scale production adoption.

Thanks to integrated autoscaling, an MCP server can automatically adjust the number of its replicas based on CPU and memory load. Some companies are already reporting stable operations with hundreds of agents connected simultaneously to dozens of servers, without loss of performance or degradation of response time.

Security and Governance: Essential Safeguards

The scalability of the Model Context Protocol (MCP) opens up considerable prospects for businesses — but it would be dangerous without a rigorous security infrastructure . Every new connection, every server added, every exposed capability multiplies the potential attack surface. This is why organizations deploying MCP at scale implement multi-layered security mechanisms , inspired by web standards and IT governance best practices.

1. OAuth 2.1 Authentication: The First Line of Defense

MCP servers are now considered true "OAuth Resource Servers," on par with critical enterprise APIs. Each access token is explicitly linked to a server via the RFC 8707 – Resource Indicators standard, preventing token mis-redemption attacks, where a token valid for service A would be fraudulently reused on service B.

This granularity of authentication ensures that:

Each AI agent acts under a traceable and verified identity ,
Servers only communicate with authorized clients,
And any attempt to reuse a token outside its perimeter is automatically blocked.

In practice, large companies are now integrating their MCP servers into their centralized Identity Infrastructure (IAM) , ensuring total consistency between human access management and software agent management.

2. RBAC and Granular Permissions: Security Closer to the Business

Role-Based Access Control (RBAC) is emerging as the preferred model for managing permissions in an agentic environment. Each MCP server defines finely segmented access scopes :

Read-only or read/write,
Access to a limited perimeter of resources (a database, a folder, a project),
Explicit prohibition of certain critical actions.

This granularity allows the "principle of least privilege" to be applied to each agent. A customer support agent, for example, can query assistance tickets in Zendesk but will have no visibility on financial data in PostgreSQL .

Thus, security becomes not a global lock, but a contextualized web of permissions , adapted to the role of each agent and the sensitivity of the data handled.

3. Human-in-the-Loop: Supervision in Critical Workflows

In high-risk environments — finance , health, industrial production, defense — the principle of human supervision remains essential. MCP facilitates this integration via mandatory validation points in workflows:

When a financial transaction exceeds a certain threshold,
When an agent attempts to modify a production environment ,
Or when an irreversible action is detected,
explicit human approval is required before execution.

This "Human-in-the-loop" approach combines the speed of AI execution with human prudence and discernment. It prevents automated drifts while strengthening trust in agentic AI systems.

Internal Registry and Governance: Preventing Shadow AI

Finally, companies deploying MCP at scale implement an internal governance registry . This registry lists all authorized MCP servers, their versions , their capabilities , and their life cycle :

Controlled activation / deactivation ,
Update history ,
Verification of signatures and dependencies ,
Automatic uninstallation of obsolete servers .

This approach helps to avoid the proliferation of unapproved tools — the infamous Shadow AI — where unvalidated connectors could access sensitive data without supervision.

As organizations adopt agentic mesh architectures , this type of registry becomes the equivalent of an Active Directory for AI : a central system of reference that guarantees the coherence, security, and compliance of interactions between agents.

Comparison with Existing Multi-Agent Architectures

The Model Context Protocol (MCP) is not intended to replace agent orchestration frameworks like AutoGen , LangGraph , or CrewAI . It integrates with them — and, above all, strengthens them. Where these frameworks orchestrate the logical collaboration between agents, MCP provides the universal interoperability layer that allows them to act in the real world , by accessing external tools, data, and services in a standardized and secure manner.

In other words: the frameworks organize the collective thought of the agents , while MCP gives them the hands and eyes to act.

AutoGen (Microsoft): Conversational Collaboration

Developed by Microsoft, AutoGen is designed to create agents capable of dialoguing with each other and collaborating through dynamic message exchanges. Each agent can define its role, negotiate its contribution, and adapt its behavior according to the context — much like a project team that self-organizes based on emerging needs.

Role of MCP in AutoGen: AutoGen uses MCP to expose external tools accessible to its agents. Thanks to the autogen_ext.tools.mcp module, agents can interact with any MCP server via STDIO or SSE transports.

Concrete Example:

A search agent queries a GitHub MCP server to retrieve relevant code.
It then transmits the information to a manager agent , which creates a ticket in Jira via a dedicated MCP server .
All without specific integration code: each action relies on the standard protocol.

Typical Use Case: A multi-source search system where one agent simultaneously queries GitHub , Jira , and Confluence via MCP, while a second agent synthesizes the results and drafts a collaborative summary note.

LangGraph (LangChain): Stateful and Branched Workflows

Originating from the LangChain ecosystem , LangGraph is a framework designed to create complex graph workflows , where the agent's decisions depend on context, conditions, and intermediate feedback. It excels in non-linear processes , capable of introducing loops, branches, and backtracking.

Role of MCP in LangGraph: The stateful model of MCP naturally complements the explicit state management logic of LangGraph.

Two modes of integration exist:

LangGraph can expose its own agents as MCP servers , accessible to other frameworks.
Conversely, it can consume MCP servers as tools in its workflows, thus leveraging external data sources or services.

Typical Use Case: A documentary research assistant that, via MCP, accesses vector databases and business APIs, while LangGraph orchestrates query refinement loops and consistency checking .

This combination makes it possible to design AI systems capable of reasoning, testing, correcting, and retrying — with an unprecedented level of autonomy.

CrewAI: Teams of Specialized Agents

CrewAI stands out for its organizational approach: it divides agents into structured teams , each with its roles, objectives, and specialties. A drafting agent, a research agent, and a revising agent can collaborate on a single deliverable, all while sharing tools and data.

Role of MCP in CrewAI: MCP servers here become resources shared among team members . Each agent can invoke the same tools via MCP, ensuring uniform coherence and security.

Typical Use Case: In a cybersecurity scenario, a Recon Agent uses an nmap MCP server to scan a network. The results are transmitted to an Intel Analyst Agent , who analyzes them using another MCP server connected to a threat database. Finally, a Reporting Agent compiles everything into a report for human teams.

Comparative Table: MCP vs. Orchestration Frameworks

Nature — MCP: AI-tool connection protocol | AutoGen: Conversational multi-agent framework | LangGraph: Graph-based workflow orchestrator | CrewAI: Agent team orchestrator
Focus — MCP: Standardizing access to data/tools | AutoGen: Dynamic AI-to-AI collaboration | LangGraph: Stateful workflows with branching | CrewAI: Agents with defined roles and goals
Statefulness — MCP: Contextual sessions | AutoGen: Conversational memory | LangGraph: Checkpoints and explicit state management | CrewAI: Shared state between agents
Interoperability — MCP: Universal (all LLMs and frameworks) | AutoGen: Compatible with MCP via modules | LangGraph: Compatible with MCP | CrewAI: Compatible with MCP
Typical Use Case — MCP: Connecting agents to CRMs, databases, APIs | AutoGen: Agent negotiation, multi-agent code review | LangGraph: Complex analytical pipelines with conditions | CrewAI: Projects requiring specialization (writing, research, QA)

In practice: Companies often combine MCP with an orchestration framework.

For example, a system might use LangGraph to orchestrate the business logic, MCP to connect agents to real systems (Salesforce, PostgreSQL, AWS), and AutoGen to manage interactions between specialized agents.

Advanced Use Cases: From Document Search to Multi-Contextual Automation

The Model Context Protocol (MCP) is not just a technical standard: it is already demonstrating its impact in concrete deployments, for developers, AI engineers, and business teams alike.

Intelligent Document Search

The Documentation Search MCP server perfectly illustrates the added value of the protocol. Designed to aggregate semantic search over more than 100 documentation sources (LangChain, LlamaIndex, OpenAI, AWS, Hugging Face, etc.), it offers AI agents a fluid and contextual access capability to the entirety of contemporary technical knowledge.

Thanks to the MCP protocol, an AI agent can:

Dynamically discover the server's capabilities — for example: get_docs , semantic_search , get_learning_path .
Query multiple documentations simultaneously from a single request, such as: "Compare state management in React and Vue".
Provide contextualized code examples and structured learning paths , tailored to the developer's level and objective.

The impact is immediate: developers and AI engineers no longer need to switch between web pages, consoles, and IDEs. Their technical assistant consults official documentation directly from the development environment , accelerating understanding and drastically reducing the cognitive load associated with context switching.

Measured Benefit: a productivity gain of 40 to 60% on monitoring, debugging, and learning new frameworks tasks.

Specialized Business Assistants

Beyond technical uses, the MCP is establishing itself as a foundation for business automation for companies. It allows for the creation of specialized agents capable of navigating between multiple tools, analyzing the business context, and executing end-to-end actions, all while respecting internal governance and security rules.

Finance: Intelligent Customer Support Management

Agents connected via MCP to the CRM and ticketing tools analyze incoming tickets in real-time. They assess urgency based on content, cross-reference customer data, and automatically create priority tasks in project management tools (Jira, Linear, Monday.com). Result: a significant reduction in average resolution time and better allocation of support resources.

Healthcare: Patient Planning and Coordination

Medical planning assistants, connected to multiple systems via HIPAA-compliant MCP servers , access patient records , medical calendars , and practitioner availability . They automatically suggest optimized slots based on patient constraints and compliance requirements. All under human supervision , with complete traceability of decisions.

E-commerce: After-Sales Chain Automation

AI agents manage refund requests , verify real-time inventory (via an MCP server connected to the ERP), and coordinate logistics with carriers — without manual intervention. This type of multi-system orchestration reduces the average processing time for complex orders and improves customer satisfaction.

Case Study: Goose, Block's (formerly Square) Unified Agent

At Block , the design , product , and support teams use an internal agent named Goose , based on the MCP protocol.

Goose acts as an intelligent gateway between several internal tools:

it automatically generates product documentation ,
processes and classifies support tickets ,
and helps prototype new features.

Thanks to the modular approach of MCP, Goose combines several servers — Slack , Snowflake , Jira , Google Drive — within a unified interface . The results are spectacular: a 75% reduction in time spent on recurrent engineering tasks , and better collaboration between technical and non-technical professions.

Multi-Contextual Automation: Integrated Workflow

One of the most telling use cases for understanding the modular philosophy of the Model Context Protocol (MCP) is that of automatic weekly meal planning . Behind an apparent simplicity lies a complete demonstration of what a composable architecture based on interconnected specialized servers allows.

1. A simple workflow, orchestrated by a single agent

The user starts by formulating a natural intention: "Prepare me an Italian meal plan for next week."

Based on this instruction, several MCP servers collaborate:

Context Selection — The user chooses a cuisine type ("Italian") via an MCP prompt exposed by the agent. This choice determines the culinary context and constraints (preparation time, diet, budget, etc.).
Meal Plan Generation — The Recipe MCP server is invoked. It compiles a complete weekly plan , with detailed recipes, cooking times, and nutritional recommendations.
Shopping List Creation — A second MCP server converts the recipes into a structured shopping list (by aisle, by quantity, or by supplier).
Final Execution — Finally, an action server takes over: printing via a thermal printer , automatic sending by email , or direct publication to a collaborative tool like Notion or Slack .

The whole process executes fluidly, without any step being manually coded — each module declares and connects dynamically via the protocol.

2. A modular, scalable, and replaceable architecture

This scenario shows that each server — recipes , shopping list , printing or distribution — is autonomous and interchangeable . A server can be replaced, improved, or reused without altering the rest of the workflow.

the recipe server can be replaced by another, specializing in vegetarian or dietary cuisine;
the printing server can be replaced by a Google Sheets , Trello , or internal ERP integration;
the central agent remains identical — it orchestrates without depending on a specific implementation.

This logic embodies the plug-and-play philosophy of MCP : each server provides a capability, each agent composes these bricks to meet a complex intention, and the whole remains stable, traceable, and extensible.

3. Transposition to the Enterprise: On-Demand Business Workflows

This same pattern applies perfectly to professional environments. In an organization, an MCP agent can orchestrate a complete sequence of tasks, such as:

the automatic generation of a report from an internal database;
the sending of a contextualized notification on Slack;
the updating of a Power BI dashboard ;
followed by the archiving of deliverables in SharePoint or Google Drive.

Each step is handled by a specialized MCP server , with human supervision possible at any time. Thus, the company can compose its own digital value chains — reusable, audited, and adaptable — without depending on a closed architecture or a single AI provider.

4. In Synthesis: From Recipe to Strategy

This culinary example clearly illustrates the founding principle of the Model Context Protocol: separating the capabilities of agents from the integration code, to make AI truly modular, governable, and scalable.

What works for a meal plan works equally well for a management report , an R&D analysis , or a regulatory compliance procedure . In all cases, the MCP acts as an invisible backbone connecting the intelligent bricks of a system — an architecture where each server is a competence module , and each agent becomes an autonomous conductor .

GitOps and Software Development

The integration of the Model Context Protocol (MCP) into Integrated Development Environments (IDEs) marks a decisive step in the convergence between software engineering and agentic artificial intelligence . The main editors — JetBrains (AI Assistant) , Cursor , Visual Studio Code , and Replit — now integrate the protocol natively , allowing their AI assistants to access the real-time project context , Git repository, and associated tools.

This evolution transforms the code assistant into a true technical collaborator , capable of interacting dynamically with the developer's environment via specialized MCP servers .

Automated Code Review: From Syntax Analysis to Contextual Understanding

Thanks to MCP, agents integrated into IDEs can directly access Git repositories and recent code changes (diffs) . The assistant no longer acts as a simple syntax corrector: it analyzes business logic , detects inconsistencies, identifies style drifts , and suggests targeted refactorings .

Example:

An agent connects to a GitHub or GitLab MCP server ,
retrieves the diffs of a pull request ,
executes a code analysis based on internal rules,
then automatically comments on problematic segments with explainable suggestions.

Benefit: significant time savings for QA and DevOps teams, reduction in the error rate in production, and standardization of development practices.

Legacy Code Migration: AI-Assisted Refactoring

Migrating a legacy codebase (e.g., from Python 2 to Python 3, or from an obsolete framework to a modern architecture) often represents a colossal effort. With MCP, this task becomes progressive, guided, and documented .

The AI agent can:

scan the existing code via an MCP Filesystem or Git server ,
automatically identify obsolete patterns or incompatibilities,
consult the most recent documentation of the frameworks concerned (via an MCP Docs server),
and propose refactorings compliant with current standards .

Suggestions can be submitted as merge requests or applied locally under human supervision.

Benefit: 60 to 80% reduction in migration time and rapid homogenization of practices across large teams.

Unit Test Generation: Automating Coverage and Compliance

MCP-based agents also allow for the automation of generating missing unit tests — a task often tedious but critical for software quality.

Typical operation:

the agent queries the documentation of the testing framework (via a dedicated MCP server, e.g., pytest_mcp or jest_mcp ),
identifies uncovered functions in the project,
and automatically generates tests adapted to the codebase conventions.

Developers can then validate, adjust, or execute these tests directly from the IDE.

Benefit: rapid increase in coverage rate, improvement in deployment reliability, and reduction of regressions over the long term.

Towards the Augmented IDE: Contextualization, Collaboration, and Autonomy

These new uses powered by the MCP transform the IDE into a complete agentic environment :

the project context is understood and exploitable by agents;
documentation, repositories, and frameworks become accessible in real-time;
interactions are standardized, auditable, and secured .

The assistant is no longer a simple input aid tool: it becomes a cognitive partner , capable of acting, reasoning, and learning in the same space as the developer.

The result? A fusion between development and contextual intelligence , where code, tools, and agents are part of the same creative flow — that of a fluid and explainable human-machine collaboration .

Perspectives: Towards an Inevitable Standard in 2026

The adoption of the Model Context Protocol (MCP) is experiencing spectacular growth, comparable to that of the major protocols in the history of the Internet.

In less than a year, MCP has moved from an experiment initiated by Anthropic to a key infrastructure of the AI ecosystem , supported by a self-reinforcing network dynamic.

Accelerated Adoption and Network Effects

Recent adoption data confirms the speed of the protocol's expansion:

More than 5,500 MCP servers are now listed on public registries (October 2025).
The 20 most popular servers alone generate over 180,000 monthly searches .
80% of them are deployed in remote mode , a sign of adoption in production in cloud and multi-agent environments.
Global ecosystem usage is growing at an average monthly rate of +33% , supported by the arrival of new frameworks and SDKs in all major programming languages.

This trajectory recalls the early days of the Web: the more sites adopt HTTP , the more logical it becomes to make it the universal communication standard. Similarly, the more MCP-compatible agents there are , the more profitable it becomes for SaaS editors and companies to implement an MCP server — thus creating a virtuous growth loop .

The ecosystem is now expanding in all directions: from DevOps to scientific research, from e-commerce to digital health. MCP is gradually becoming the invisible backbone of connected AI .

The network effects of the MCP rely on a simple principle: each new server, each new compatible agent increases the value of the entire system.

Direct Effect: a new MCP-compatible AI agent can immediately interact with hundreds of existing servers (GitHub, Jira, Snowflake, Notion, etc.).
Indirect Effect: the more servers multiply, the more useful agents become — and the more developers are encouraged to join the ecosystem.

This dynamic leads to structural convergence : the MCP is no longer an implementation choice, but an interoperability prerequisite , exactly like HTTP for the Web or TCP/IP for networks .

Major platforms are starting to adopt this paradigm: OpenAI, Microsoft, Google, Anthropic, and AWS all recognize the need for a neutral and universal protocol to orchestrate interactions between AI, tools, and enterprise systems.

Roadmap 2026: Standardization and Maturation

The development of the protocol follows a clear and ambitious trajectory, structured around five priority axes:

Complete Standardization by 2026 — Publication of a stable specification, compliance frameworks, and official certification of implementations to ensure long-term interoperability.
Native Multimodality — Support for video , audio , and real-time streams through streaming and chunking. Objective: to allow multimodal agents (Claude, Gemini, GPT-5) to interact fluidly with their sensory environments.
Enhanced Security — Mandatory switch to OAuth 2.1 for all connections, prevention of token mis-redemption (via RFC 8707), integration of Decentralized Identities (W3C DID) for traceability and confidentiality.
Centralized Registry — Creation of a true MCP App Store , allowing for automated discovery , versioning , verification , and community rating of servers.
Inter-Agent Interoperability (A2A) — Integration of Google's Agent-to-Agent (A2A) protocol for direct communication between agents, thus complementing the MCP, which is focused on agent-tool communication.

According to market projections (Gartner, CB Insights, McKinsey Digital), the MCP ecosystem could reach $10.3 billion in value in 2025, with an estimated Compound Annual Growth Rate (CAGR) of 34.6% .

This figure includes:

MCP server solutions,
marketplaces and registries,
integrated orchestration frameworks,
and associated governance and security services.

By 2026 , the MCP is expected to establish itself as the de facto standard for integration between AI and business tools — the equivalent of HTTP for applied artificial intelligence .

Challenges to Overcome

Despite the enthusiasm generated by the Model Context Protocol (MCP) , several challenges remain before fully generalized adoption. Like any emerging technology destined to become a standard, the MCP must still cross stages of technical, security, and cultural maturation .

Maturity of Technical Specifications

While the foundations of the protocol are solid, some dimensions still require iteration:

bandwidth management for very large-scale deployments,
optimization of session maintenance between thousands of simultaneous agents,
and standardization of performance metrics to ensure interoperability between implementations.

These challenges are comparable to those faced by HTTP or Kubernetes in their early days: a necessary adjustment phase before the standard stabilizes.

Security and Compliance in Production

Companies operating in regulated environments (banking, healthcare, energy) are proceeding cautiously. As long as the security practices and compliance frameworks of the MCP are not fully proven, some organizations hesitate to entrust autonomous agents with direct access to their critical systems.

Current work on the generalization of OAuth 2.1 , the management of Decentralized Identities (DID) , and mandatory human supervision should remove these obstacles within the next 12 to 18 months.

Training and Organizational Culture

Adopting the MCP is not limited to installing a protocol: it is a paradigm shift . Teams must learn to think in agentic architectures , compose specialized servers , and integrate AI governance into their development cycle.

This skills upgrade requires structured support — training, documentation, feedback — and an evolution of roles within organizations: data engineers become cognitive interface architects , and developers agent orchestrators .

The pioneers show the way: Companies like Block , Apollo GraphQL , or Rocket Companies have demonstrated that these obstacles are surmountable through a progressive approach: targeted pilot projects on high-impact use cases, intensive training of technical and product teams, and implementation of strict governance from the first experiments.

Conclusion

The Model Context Protocol fundamentally transforms the way AI systems interact with their environment. By replacing monolithic and proprietary integrations with a standardized, modular, and observable agentic mesh , MCP lays the foundations for an open, governable, and sustainable infrastructure for artificial intelligence.

This protocol redefines the flows between agents, APIs, and humans , bringing tangible benefits:

drastic reduction in integration costs ,
development time divided by two ,
total auditability of agentic decisions,
and horizontal scalability adapted to production deployments.

A Catalyst, Not a Substitute

Compared to multi-agent frameworks such as AutoGen , LangGraph , or CrewAI , the MCP does not replace them — it potentiates them. It provides a unified and secure access layer to tools, data, and services, allowing these frameworks to focus on the logic of orchestration, coordination, and collaboration between agents.

The use cases already operational — from intelligent document search to multi-contextual business automation — prove that the protocol is no longer experimental: it is ready for critical environments .

The Next Standard for AI Integration

With rapidly accelerating adoption , an ecosystem of over 5,500 servers , and massive support from tech giants , the MCP is following the trajectory of major standards in digital history. Just as HTTP unified the Web, MCP is set to unify the AI ecosystem.

For organizations in transformation, mastering the MCP is no longer an option — it is a strategic imperative for building governable, scalable, and sustainable artificial intelligence systems.

AI as a cognitive exoskeleton: DigitalKin and Luc Julia on the future of human expertise

m.boisis@digitalkin.ai — Thu, 25 Sep 2025 09:29:33 GMT

At a Medef Lyon Rhône roundtable, DigitalKin's cofounder Emmanuel Théry and AI pioneer Luc Julia debated whether AI will replace human expertise or amplify it.

Will AI multiply our potential or make human expertise obsolete? That was the question at "Ça Match," a roundtable hosted by Medef Lyon Rhône at La Sucrière on September 16, 2025.

On stage: Luc Julia, Franco-American AI specialist widely regarded as one of the grandfathers of Siri, now Chief Scientist at Renault. And Emmanuel Théry, cofounder of DigitalKin, the Lyon-based startup building autonomous AI agents for high-stakes professional work.

"AI is our augmented intelligence"

For Luc Julia, the answer is clear: humans create AI and use it. Not the other way around. He sees cooperation, not competition. AI helps us do certain things better, or do things we simply couldn't do before. But it remains specialized, task-focused, and fundamentally dependent on human direction.

"AIs work with what we already know, on existing data. They don't invent anything. They follow our prompts and go searching for ideas that already exist. AI is, in fact, our augmented intelligence," Julia explained.

"AI becomes a cognitive exoskeleton"

Emmanuel Théry agreed, and pushed the idea further. DigitalKin builds what he calls "digital colleagues" for expert professionals, particularly in R&D. These are not generic chatbots. They are specialized multi-agent systems that integrate each expert's specific methodology, knowledge that is not publicly available.

"An expert always has a precise way of handling a subject. It's not ChatGPT that can help, you need specialized agents," Théry said. "Our Kins are multi-agent systems that seek the best of each model and integrate human expertise that isn't public. The expert is fast but limited by time. AI is fast and frees the expert for other tasks. AI becomes a cognitive exoskeleton."

The ROI problem, and why most companies get it wrong

Both speakers addressed a striking reality: while 80% of companies have tested AI and 40% have deployed it, only 5% report a clear return on investment, according to a recent MIT study. The reason, they argued, is not that AI doesn't work. It's that most deployments are not deeply integrated into the company's actual processes and context.

The competitive advantage of AI doesn't lie in raw productivity gains. It lies in a true symbiosis between AI and human expertise, where each amplifies the other's potential.

Creative destruction, not job destruction

As for the fear that AI will destroy jobs, Luc Julia was direct: "We don't know anything yet. But for me, we're in a typical period of creative destruction that comes with every major technology shift, and it ends with a net positive balance."

Emmanuel Théry and Luc Julia spoke at "Ça Match," organized by Medef Lyon Rhône at La Sucrière, Lyon, on September 16, 2025. Originally covered by BrefEco.

The French RTC – State-of-the-Art Kin writes your state of the art in record time

Tue, 16 Sep 2025 10:27:21 GMT

Researchers, R&D managers, CIR experts, Research and Innovation Funding Consultants: have you already found yourself facing these challenges?

Endless hours spent justifying your research projects, in particular, building Literature Reviews In Your CIR Technical Files
Technical states of the art that seem complex to you to realize: A specialized niche subject? Hard to find scientific publications?
Funding that sometimes escapes you for lack of a solid, well-sourced, exhaustive argument.

These obstacles hinder innovation and increase the burden on R&D teams of all companies combined.

‍

But imagine a world where...

The state of the art of your CIR files are prepared in a few minutes rather than weeks or months.

Your CIR literature reviews Are comprehensive, well-sourced, accurate, and up to date

Your financing acceptance rates are more secure thanks to very high level declarations

You can focus your efforts on Innovation and research , not on the administrative files that accompany the financing of your work

At DigitalKin , we designed the business solution - The Kin Research Funding - CIR - Who Makes You Truly Save Time and Improve Significantly The quality of your CIR state of the art.

‍

How?

Through the unique combination of human expertise in writing state of the art for the financing of research and the power of autonomous and multi-model artificial intelligence. ‍

‍

What is the Kin CIR? ‍

The Kin CIR is not a simple assistant or chatbot, it is a real expert AI collaborator who knows the tax requirements of public funders and ensures total compliance with their expectations in terms of technical arguments.

‍ It analyzes your R&D projects in depth: It pre-evaluates the activities eligible for the CIR accurately according to the official Frascati criteria
It Writes States of the Art Independently and puts forward a powerful scientific argument to be inserted directly into your CIR technical files

It helps you quickly find suitable scientific sources in the form of comprehensive and substantiated documentary bodies to defend or deepen your cases in the event of an audit.

‍

What differentiates us... Top 3!

From research to synthesis: Unlike traditional tools that simply list articles, the Kin État de l'Art - CIR analyzes, cross-references and synthesizes information into a written report that is structured and directly usable - supporting quotations and bibliography for each argument put forward!
From generic summaries to expert state of the art: Our “Objective-driven” approach ensures that Kin understands the context, intent, and specific needs of each user expert for results that are truly relevant to writing your CIR files.
From dependence on a single model to multi-model intelligence: Our technology selects the best language model for each task and assures you of extremely qualitative results regardless of your field of research.

Give it a try! Get a state of the art written by Kin on your priority research topic

Preserving what makes companies unique with agentic AI

Mon, 15 Sep 2025 21:22:11 GMT

At the MEDEF "Ça Match" event in Lyon, DigitalKin's CEO presented how Kins protect and amplify the singularity of each organization, in a world where generic AI tends to standardize.

On September 16, 2025, the MEDEF Auvergne-Rhône-Alpes gathered 2,000 business leaders at La Sucrière in Lyon for "Ça Match," its flagship innovation event. Among the speakers: Emmanuel Théry, CEO and cofounder of DigitalKin, the Lyon-based startup pioneering agentic AI for high-stakes professional work.

The standardization risk

As AI adoption accelerates, a paradox is emerging. The more companies rely on the same generic AI tools, the more their outputs, analyses and decisions start to look alike. For organizations whose competitive advantage lies in proprietary methods, specialized knowledge or unique expertise, this convergence is a strategic threat.

"The question is no longer whether AI will transform our businesses, but how to prevent it from making them all identical," Théry told the audience.

Kins: AI that amplifies singularity instead of erasing it

DigitalKin's approach stands in direct contrast to generic AI. Its Kins are multi-agent systems designed to reproduce and amplify the methods specific to each expert and each organization. Rather than applying a one-size-fits-all model, every Kin is configured with the client's own ontology: their vocabulary, their methodology, their decision criteria, their quality standards.

"Our Kins don't replace human expertise. They multiply it," Théry explained. "A Kin without real expertise behind it is just another chatbot. A Kin carrying your methodology becomes something no competitor can replicate."

Already adopted by leading organizations

DigitalKin counts among its early clients companies such as Boiron laboratories and a major household appliance manufacturer. These organizations report a significant multiplication of their capacity to produce complex strategic documents, while maintaining or improving their usual quality standards.

A humanist vision of AI

A winner of the French Tech Seed program and member of NVIDIA Inception, DigitalKin asserts an exacting and humanist vision of AI. In a landscape where fears of skill loss and thought standardization are growing, the company offers a clear response: build AI agents that valorize internal specificities rather than erase them.

The technology is built on three principles: full configurability by the user, complete traceability of every action, and human validation at every critical step. The result is AI that professionals can endorse and stand behind.

Emmanuel Théry spoke at "Ça Match," organized by MEDEF Auvergne-Rhône-Alpes at La Sucrière, Lyon, on September 16, 2025.

The Model Context Protocol (MCP): architecture, concepts and ecosystem

Tue, 15 Jul 2025 09:00:00 GMT

Fundamental Architecture

The Model Context Protocol (MCP) is an open standard designed to fundamentally rethink how artificial intelligence applications interact with their digital environments.

Introduced by Anthropic in November 2024 , MCP aims to unify access of language models (LLMs) to external data sources and tools, replacing fragmented and proprietary integrations with a universal, secure, and structured interface.

Where businesses previously had to develop a separate connector for each model and tool combination (creating a type complexity N×M ), MCP transforms this problem into a much more manageable architecture, such as M+N : each model implements a single client , each tool a single server , and both communicate according to a common grammar.

MCP Host

The MCP architecture is based on a standardized client—server model , organized around three main components that work together to provide rich context to language models:

The creation and supervision of MCP clients
The permission management and user permissions
The application of security policies and isolation rules
The coordination of exchanges between LLMs and external servers

MCP Client

The client is the active bridge between the host and the MCP servers. It maintains a 1:1 isolated connection with each server and ensures the consistency of exchanges via a persistent session with status.

Its responsibilities include:

The negotiation of the protocol and the discovery of capabilities from the server
The bi-directional message management and events
The preservation of security borders between servers
And the maintenance of subscriptions and notifications , guaranteeing smooth and continuous communication.

MCP Server

The MCP server is a lightweight program dedicated to a specific function.

It exposes its capabilities — resources, tools, prompts or sampling requests — according to standardized protocol primitives.

A server can function independently , locally (isolated process) or remotely (HTTP/SSE service). Each server stays responsible for a specific functional perimeter : it can provide access to a database, a business API, a file system, or a scientific computing environment.

Thanks to this approach modular and decoupled , the same host can connect to several MCP servers in parallel , each bringing a different skill or type of resource, while maintaining insulation, safety and traceability exchanges between them.

Protocol Structure and JSON-RPC Messages

The Model Context Protocol (MCP) relies on JSON-RPC 2.0 as the technical foundation for all exchanges between clients and servers. This choice is based on a logic of simplicity and universality: JSON-RPC provides a message format that is lightweight, human-readable, language-agnostic, and already widely adopted in the software ecosystem.

The Three Fundamental Message Types

1. Requests
‍ These are the initial messages sent to execute an operation.Each request contains:

a unique identifier (string or integer, never null);
a method to invoke, such as tools/call or resources/read ;
and optional parameters .The ID is used to link each request to its response and must never be reused within a given session.

2. Responses
Responses reuse the ID of the corresponding request and include either a result field (success) or an error field (failure)—but never both. Error codes are standardized as integers and allow for precise diagnosis of incidents.

3. Notifications Unidirectional messages that do not require a response, notifications inform of a state change or an event without blocking the main execution flow. They are commonly used to signal resource updates, session modifications, or the arrival of new real-time events.

Lifecycle Management and Capability Negotiation

MCP defines a strict lifecycle management for connections between clients and servers to ensure session stability and shared state coherence. This cycle takes place in three distinct phases :

1. Initialization Phase

Upon connection establishment, the client sends an initialize request to the server. This step allows:

negotiating the protocol version ;
exchanging the supported capabilities on both sides (available tools, accessible resources, message formats).
The server responds by specifying its own capabilities, after which the client emits an initialized notification to confirm that the session is operational. This phase lays the groundwork for a contractual and verifiable communication .

2. Operation Phase

Once the session is established, exchanges become bidirectional: requests, responses, and notifications flow according to the negotiated capabilities.

It is during this active phase that:

tools are invoked;
resources are read or written;
and prompts , requests , or contextual actions are executed.
Each interaction respects the authorization, audit, and isolation logic specific to the session.

3. Shutdown Phase

When the session is no longer needed, closure is gracefully initiated by the client or the server. This final step releases allocated resources, properly closes open channels, and prevents the appearance of orphaned connections that could unnecessarily consume resources or create security vulnerabilities.

Capability Negotiation

One of MCP's strengths lies in its dynamic discovery and negotiation mechanism .

Upon initialization, each participant announces:

the methods it supports;
the data formats it accepts;
and the extensions it supports.

‍

The client and server thus build an explicit session contract , ensuring that only mutually compatible functionalities are used. This process prevents runtime errors, improves resilience, and makes each session predictable, traceable, and extensible .

Transport Layer: STDIO, HTTP, and SSE

The Model Context Protocol (MCP) was designed to adapt to a wide variety of environments and integration constraints.

For this purpose, it supports several transport mechanisms for communication between clients and servers, each having specific advantages depending on the usage scenario.

STDIO (Standard Input/Output) — the quintessential local transport

The STDIO mode relies on simple and direct operation: the client launches the MCP server as a child process and communicates with it via the standard stdin and stdout streams.

Each message is delimited by a newline and formatted according to the JSON-RPC 2.0 specification.

‍

This mode has several major advantages:

Extremely low latency , with no network passage.
Minimal configuration , ideal for development environments or embedded integrations.
Enhanced security , as no network exposure is necessary: communication remains confined to the local context.
1:1 relationship between client and server , ensuring total isolation of each session.

‍

It is the preferred transport for:

local integrations in applications such as Claude Desktop or VS Code ;
Command Line Interface (CLI) tools ;
and sensitive operations requiring strict control over data flows.

HTTP with Streamable — the recommended modern network transport

The HTTP with Streamable transport transforms the MCP server into a web service accessible via URL .

Unlike the STDIO model, where each client has its own process, an HTTP server can manage multiple clients simultaneously , fostering scalability and distributed deployment.

‍

The Streamable HTTP protocol introduces full bidirectional communication between client and server, while supporting:

streaming responses , allowing a continuous flow of information (useful for tracking long tasks);
concurrent management of multiple active sessions;
and fluid integration with existing web infrastructures (API Gateway, reverse proxy, load balancer).

‍

This mode is recommended for:

remote or cloud deployments ;
multi-client architectures (multiple agents connected to the same service);
and network applications that require centralized supervision or orchestration.

SSE (Server-Sent Events) — inherited compatibility

Transport via Server-Sent Events (SSE) is historically the first method used by MCP for HTTP communication.

It allows unidirectional streaming from the server to the client , but does not support full bidirectional exchange.

‍

While still maintained for backward compatibility reasons , SSE is now deprecated for new projects. Developers are encouraged to adopt HTTP Streamable , which offers greater flexibility, performance, and security.

Choice of Transport: a Question of Context

STDIO — perfectly suited for local integrations, CLI tools, and environments where simplicity and security are paramount.
HTTP Streamable — essential for network architectures, multi-agent environments, and large-scale collaborative uses.
SSE — remains useful for maintaining compatibility with older clients, but no longer recommended for modern deployments.

The Three Fundamental Primitives: Resources, Prompts, and Tools

At the heart of the Model Context Protocol (MCP) are three essential primitives — Resources, Prompts, and Tools — that allow servers to offer rich, explainable, and contextualized interactions to language models.

Resources — the informational foundation

Resources represent all the data that a model can access, whether local, remote, or dynamically generated.

Each resource is identified by a URI (Uniform Resource Identifier) , ensuring a clear structure and standardized access, regardless of the content type.

‍

Examples of URI formats:

file:///home/user/projects/api/README.md — local file;
postgres://localhost/customers/table/users — table in a PostgreSQL database;
weather://current/san-francisco — dynamic resource linked to a weather API.

‍

Servers expose these resources via the resources/list endpoint, while hosts can request them to be read using resources/read.

Each resource is accompanied by descriptive metadata (MIME type, description, update date, access rights) allowing users and models to understand its content and context before use.

Resources can be:

static , like a document or versioned source code;
or dynamic , like the latest entries in a system log or an IoT stream.

‍

Thus, the LLM is no longer confined to its training data: it accesses a universe of living, readable, and traceable information .

Prompts — encapsulated expertise

Prompts constitute the second key primitive of the protocol.

They offer reusable instruction templates encapsulating the expertise of a domain or the logic of a complex task.

Each prompt acts as an operational guide that the model can invoke with precise parameters, ensuring consistency, reliability, and homogeneity of responses.

‍

Typical examples:

A database server can offer:
analyze-schema (analyze a schema),
debug-slow-query (diagnose a slow query),
generate-migration (create a migration script).
A Kubernetes server can offer prompts to troubleshoot a cluster or analyze a deployment .
A code review server can expose prompts adapted to the team's internal style guides .

‍

Prompts can accept dynamic arguments : for example, analyze-table takes the name of a table as input, then returns a structured diagnosis including schema, indexes, and foreign key relationships .

‍

In summary, prompts constitute the protocol's business memory —a way to industrialize human expertise in the form of standardized and exploitable instructions for models.

Tools — the interface with the real world

Finally, Tools are the executable functions that a model can call to perform concrete actions in its environment.

‍

Each tool is defined by:

an explicit name ,
a clear description of its usage ,
a JSON schema detailing the input parameters (name, type, constraints, default values),
and a handler —the code that actually executes the business logic.

‍

Invocation is done via the tools/call method, where the client provides the tool name and the expected arguments.

The server executes the function, then returns the structured result (success or error).

‍

These tools allow the LLM to move beyond simple textual generation to interact operationally: execute commands, modify a database, launch a simulation, or orchestrate other services.

A Rich and Coherent Ecosystem

By combining these three primitives :

Resources for context ,
Prompts for guidance ,
Tools for action

‍

MCP offers a complete interface between the model's thought and the digital world .

‍

This tripartition transforms the language model into a connected cognitive actor , capable of reading, understanding, reasoning, and acting in a governable manner, within an ecosystem that is standardized, extensible, and traceable .

Advanced Features: Sampling and Elicitation

Sampling: the inversion of the decisional flow

The sampling mechanism marks a major evolution in the protocol's interaction philosophy.

Traditionally, it is the clients that query the language models. With sampling, MCP servers can themselves request LLM completions , thus inverting the classic communication flow.

‍

This mechanism is used when a server needs to:

make a complex decision involving reasoning
generate a response beyond deterministic rules;
or understand a semantic context that only a language model can interpret.

‍

The execution flow follows a rigorously controlled sequence:

The server sends a sampling/createMessage request to the client, containing the conversation messages , model preferences (if any), system instructions , and the required context .
The client receives this request, validates it , and displays to the user what the server wants to transmit to the model.
→ Human Checkpoint n°1 : the user can edit , approve , or reject the proposed prompt.
If the prompt is approved, the client sends it to the LLM, receives the completion , and presents it to the user again.
→ Human Checkpoint n°2 : the user can correct , validate , or refuse the generated response.
Upon approval, the final completion is returned to the server , which can then continue its processing.

‍

This dual human control guarantees total traceability of the reasoning , while allowing servers to leverage the power of language models without loss of governance .

Sampling thus embodies a balanced compromise between agentic autonomy and active human supervision —a fundamental principle of MCP design.

Elicitation: the interactive loop with the user

The second advanced capability, known as elicitation , allows MCP servers to request complementary information from the user when a necessary piece of data for the execution of a task is missing or ambiguous.

‍

This functionality transforms the server from a simple executor into a dialogic actor , capable of maintaining a contextual and adaptive exchange :

if a command requires a missing parameter (e.g., file name, date range, resource identifier), the server can directly query the user;
the client relays this request in a clear and validatable form;
the response is then reinjected into the execution flow.

‍

Elicitation significantly improves the robustness and usability of MCP-based systems.

Rather than silently failing on input errors, the protocol promotes fluid cooperation between human and machine, where each missing step becomes an opportunity for contextual adjustment.

Metadata Schema and Human Supervision

The Model Context Protocol (MCP) places human supervision at the heart of its design.

Rather than considering governance as an afterthought layer, the protocol integrates it by design : no sensitive action—whether it is data access, code execution, or a request to a language model—can be undertaken without explicit user consent .

‍

This approach guarantees that final control always belongs to the human , even in environments where multiple AI agents collaborate or make autonomous decisions.

User Consent and Explicit Control

Every operation initiated by an agent must be understood and authorized by the user.

‍

MCP implementations must therefore:

provide clear and explicit interfaces allowing the visualization, review, and approval of each action;
ensure that users know which data is being shared , with which servers, and for what purpose;
and maintain the possibility to revoke an granted authorization at any time.

‍

This principle transforms consent into a conscious and traceable act , not a simple checkbox.

Confidentiality and Data Governance

MCP hosts have the obligation to obtain explicit consent before any exposure of user data to an external server.

‍

No data must be transmitted outside the environment without prior validation, and all data must be protected by:

granular access controls (based on roles, tasks, or contexts),
encryption mechanisms for exchanges and storage,
and audit logs tracing who accessed what, when, and why.

‍

Thus, MCP ensures complete traceability of data flows and prevents any drift towards opaque or uncontrolled models.

Security of Tools

Tools , capable of executing code or manipulating real resources, represent a particular risk area.

‍

The protocol therefore imposes a principle of reinforced caution :

each invocation of a tool must be explicitly approved by the user;
interfaces must clearly display what the tool is about to execute;
and hosts must isolate execution environments to prevent side effects or compromise.

‍

The objective is to ensure that the user fully understands the scope and consequences of each action before authorizing it.

Control of LLM Sampling

LLM sampling requests , which allow a server to request a completion from a language model, are subject to even stricter supervision.

‍

Users must:

explicitly validate whether sampling can be performed;
see the actual prompt that will be sent to the LLM;
and choose which parts of the response can be returned to the server.

‍

The protocol deliberately limits the server's visibility into prompts and completions to avoid any leakage of confidential context or unintentional bias.

This approach ensures that sampling remains an assisted process , never automatic.

The Five Layers of Authentication and Authorization

MCP's security model is based on a five-level architecture , designed to ensure complete end-to-end governance:

Agent Identity.
Each AI agent has a unique and traceable identity , allowing its actions and responsibilities to be tracked.
Delegator Authentication.
The human user authenticates to the system and becomes the delegator of authority for the agents they control.
Consent Delegation.
The user defines the scope of authority for each agent—that is, what actions it can undertake, on which resources, and within what temporal or contextual perimeter.
MCP Server Access.
The agent then authenticates with the MCP server , respecting the authorizations delegated by the user.
Upstream Services.
Finally, the external APIs accessed by the server must respect the agent's identity and the inherited permissions from the human delegator.

A Protocol Centered on Trust and Traceability

This hierarchical authentication and explicit consent architecture makes MCP one of the first standards to integrate human governance as a technical component .

Each interaction thus becomes verifiable, reversible, and attributable , laying the groundwork for a Trusted Agentic AI , where autonomy never excludes responsibility.

MCP Registry: Discovery and Distribution

Officially launched by Anthropic in September 2025 , the MCP Registry constitutes a structuring step in the maturation of the Model Context Protocol ecosystem.

This centralized registry acts as a public and interoperable catalog grouping all available MCP servers, thus facilitating their discovery, integration, and governance .

‍

Designed as a metaregistry , the registry stores neither source code nor binaries , but only the descriptive metadata of the MCP servers.

It positions itself as a universal entry point allowing developers, companies, and orchestration tools to reference and exploit existing servers in a standardized manner.

Main Features of the Official Registry

Authoritative Repository
The MCP Registry is the single source of truth for all officially published public MCP servers.
It guarantees the consistency, traceability, and reliability of information related to each server available in the ecosystem.
Community Ownership
Although initiated by Anthropic, the registry is now collectively managed by the open-source MCP community .
It benefits from the support of major players such as GitHub , PulseMCP , and Microsoft , who contribute to its maintenance and governance.
This collaborative approach reinforces its legitimacy as a neutral and sustainable infrastructure.
Unified Discovery
Server creators no longer need to multiply publications: a single registration in the MCP Registry is enough to make a server visible across all compatible platforms.
All consumers—applications, orchestrators, Kins, or multi-agent frameworks—can reference the same canonical data , ensuring total interoperability .
Standardized Format
Each registry entry follows the server.json format, which exhaustively describes the elements necessary for the installation and use of a server:
Unique Identity (e.g., io.github.user/server-name)
Available Packages (places to download the server, via npm , PyPI , Docker Hub , etc.)
Runtime (execution instructions, arguments, environment variables)
Metadata (description, version, capabilities, dependencies, compatibilities)
This common format simplifies automation, updating, and synchronization between environments.

An Interconnected Ecosystem

The MCP Registry is not intended to replace existing third-party registries — such as Smithery , Mastra , Glama.ai , or MCP.so .

On the contrary, it acts as a synchronization core : these platforms can replicate or enrich the metadata of the official registry, while offering their own value-added services (advanced search engines, qualitative evaluation, thematic curation, community scoring, etc.).

‍

Thus, the official registry becomes a pivot of interoperability between the protocol's base infrastructure and more specialized application layers.

It contributes to making MCP an open, traceable, and transparent ecosystem , where every server can be discovered, evaluated, and integrated with confidence — an App Store for agentics , serving standardization and digital sovereignty.

Link Between MCP and LLM Context Window

The Context Window: An Inherent Model Constraint

The context window refers to the maximum quantity of tokens (words, symbols, or text fragments) that a language model can process simultaneously.

‍

It is a physical limit , directly linked to the model's architecture and size.

For example: a model like GPT-4-turbo can process 128,000 tokens , while a lighter model may be limited to 4,000 or 8,000 tokens .
Beyond this capacity, the model can no longer reason effectively , as it only "sees" a portion of the available text.

‍

In other words, the context window defines the model's immediate memory — its ability to "remember" what is provided to it in a single request.

The Model Context Protocol: An Orchestration Layer Above

The Model Context Protocol , conversely, has nothing to do with this internal limit.

It is an orchestration protocol that standardizes the way a model accesses its external environment : databases, files, APIs, tools, or other AI agents.

‍

MCP therefore does not seek to increase the size of the context window, but to intelligently manage the flow of information that enters it.

It defines how and when to provide relevant data to an LLM, while respecting its technical constraints.

‍

Thus, MCP acts as a context conductor , ensuring that the model:

receives only the useful information at the right time ,
interacts with reliable and standardized sources ,
and is never overwhelmed by a volume of data greater than what it can process.

Dynamic Contextualization: A Key Contribution of the Protocol

One of MCP's great strengths lies in its ability to enable dynamic and selective contextualization .

Instead of sending all data at once — which would saturate the model's window — the protocol allows servers to:

selectively provide relevant fragments throughout the reasoning process,
refresh the context based on the task's status,
and maintain coherence between successive sessions.

‍

For cases where the volume of information far exceeds the model's capacity, advanced orchestration patterns can be implemented, such as the sub-context pattern :

Divide the data into subsets (chunks) ;
Create a dedicated sub-context for each portion;
Process each sub-context independently ;
Then merge the summaries to produce a consolidated synthesis.

‍

This mechanism allows for the processing of massive volumes of data while respecting the structural constraints of the LLM.

Reference Implementations and Ecosystem

Official SDKs in More Than Ten Languages

To ensure wide and cross-platform adoption, Anthropic and the MCP community maintain official Software Development Kits (SDKs) in more than ten major languages : TypeScript, Python, C\#, Go, Java, Kotlin, PHP, Ruby, Rust, and Swift.

‍

These SDKs offer:

high-level abstractions to quickly create compliant MCP servers and clients according to the specification;
modules for automatic JSON-RPC schema validation ;
and tools to manage capability negotiation , security , and state persistence .

‍

Thanks to these libraries, a developer can implement an MCP server in a few lines of code, while benefiting from native compliance with the official protocol.

Official GitHub Repository: Reference Servers for All Use Cases

The GitHub repository model-context-protocol/servers centralizes the reference implementations developed by Anthropic and the community.

These servers serve both as pedagogical models and concrete starting points for building new integrations.

‍

Among the most popular:

Everything : a versatile test server integrating prompts, resources, and tools.
Fetch : enables the retrieval and conversion of web content into MCP resources.
Filesystem : manages secure read and write operations on local file systems.
Git : provides tools for reading, searching, and manipulating Git repositories.
Memory : offers persistent memory based on knowledge graphs.
Sequential Thinking : experiments with reflexive reasoning and task chaining logics.
Time : manages timezone conversions and complex temporal operations.

‍

To this foundation is added a galaxy of third-party servers covering almost all enterprise integrations: AWS, Azure, Google Cloud, GitHub, GitLab, Slack, Jira, Salesforce, Stripe, Shopify, PostgreSQL, MongoDB, Elasticsearch , and many others.

MCP Inspector: The Essential Development Tool

The MCP Inspector has become one of the indispensable tools for developers and integrators.

It allows connection to an MCP server and inspection of its advertised capabilities — whether tools , resources , or prompts — without requiring a complete AI client.

‍

The user can thus:

visualize the metadata exposed by the server;
test the endpoints manually;
simulate requests and verify JSON-RPC responses ;
and diagnose errors before any production deployment.

‍

In other words, MCP Inspector plays a role equivalent to Postman for REST APIs or GraphiQL for GraphQL : an interactive testing environment essential for understanding, auditing, and securing MCP integrations.

‍

Thanks to this set — multilingual SDKs, reference servers, industrial integrations, and diagnostic tools — the MCP ecosystem forms a complete, open, and evolving infrastructure , supporting the vision of universal interoperability between AI agents and business environments .

Integration with Orchestration Frameworks

LangGraph + MCP: State-Aware Orchestration and Universal Interoperability

LangGraph is a framework designed to orchestrate complex AI agent workflows, relying on a graph of nodes representing decisions, transitions, and shared state between agents.

The integration of MCP into LangGraph allows each agent to consume the capabilities exposed by MCP servers , and thus to access, in a standardized format, external prompts , resources , and tools .

‍

Concretely, a LangGraph agent can:

dynamically list available MCP servers;
invoke their tools as part of its reasoning or a sub-workflow;
and maintain a persistent contextual state , ensuring the continuity of reasoning between successive calls.

‍

This hybrid architecture combines LangGraph's structured state management with the functional richness of the MCP network.

Pilot projects already demonstrate decentralized multi-server systems , where LangGraph agents orchestrate distributed workflows, delegating tool execution to remote MCP servers via SSE and STDIO transports.

‍

Result: an ecosystem where every agent can plan, reason, and act by leveraging the resources of a global contextual mesh .

CrewAI + MCP: Standardizing Access to Enterprise Tools

The CrewAI framework, specializing in the coordination of autonomous agent teams , also integrates natively with MCP via the crewai-tools library and its dedicated adapter, the MCPServerAdapter .

‍

This adapter ensures:

secure connection to MCP servers via the chosen transport (STDIO or SSE);
automatic discovery of available tools on each server;
and the creation of compatible CrewAI wrappers , exposing these tools in the framework's own language.

‍

Thus, a CrewAI agent can immediately access a standardized ecosystem of over 50 "enterprise-grade" integrations , notably via services like Klavis or PulseMCP .

This approach unifies connector management and avoids the proliferation of specific scripts for each environment.

‍

In practice, MCP acts as a universalization layer : CrewAI maintains its multi-agent coordination logic, but offloads tool and resource management to an open and governable infrastructure.

Microsoft AI Foundry + MCP: Unifying Azure AI Services

The Microsoft Azure AI Foundry platform has also adopted MCP as a standardized interface to connect its agents to external tools and services.

‍

Thanks to this integration, Foundry agents can securely connect to third-party environments, while Copilot Studio agents can invoke these capabilities mid-dialogue , without breaking context.

‍

The MCP Server for Azure AI Foundry (currently in experimental version) provides a unified integration layer between the different services of the Azure ecosystem:

model exploration and comparison ;
knowledge base and embedding management ;
performance evaluation and monitoring via natural language .

‍

This convergence between MCP and Foundry illustrates a deep-seated movement: the transformation of large cloud environments into open agentic infrastructures , capable of interoperating through a common standard.

‍

By uniting frameworks like LangGraph , CrewAI , and Azure AI Foundry around the same protocol, MCP positions itself as the universal exchange language for distributed intelligence .

It becomes the foundation for a shared technical governance , where every agent, regardless of its execution environment, can dialogue with others according to transparent, auditable, and secure rules.

Differentiation: MCP vs RAG vs Function Calling

MCP vs RAG: From Static to Living Context

Retrieval-Augmented Generation (RAG) allows a model to enrich its responses by retrieving passages from a vector knowledge base .

These bases contain pre-embedded documents , meaning they are converted into numerical representations (vectors) before querying.

The model then queries this base via a similarity search to retrieve relevant excerpts.

‍

Limitation: This approach works well for static documentary corpora (reports, articles, manuals), but becomes inefficient when data changes frequently or needs to be retrieved dynamically.

‍

The Model Context Protocol , on the other hand, operates on a different logic.

Instead of querying fixed documents, MCP allows LLMs to access structured, real-time data directly , from databases , APIs , business tools , or SaaS services .

It requires no pre-embedding , and acts on demand (runtime) to fetch the correct information, at the right time.

‍

In Summary:

RAG = useful for searching static, time-frozen corpora.
MCP = essential for living, sensitive, or evolving data , requiring a direct and secure query.

‍

MCP thus transforms a model's contextualization from simple document retrieval into a dynamic and governed process .

MCP vs Proprietary Function Calling: From Dependence to Open Standard

Function calling introduced by OpenAI (and used in other forms in ChatGPT plugins ) marked an important step in LLM connectivity.

They allow models to invoke pre-defined functions by the developer, but these approaches are proprietary : each vendor imposes its own format, its own rules, and its own connectors.

‍

Result: a strong dependence on the vendor's ecosystem.

A connector developed for GPT-4 is not compatible with Claude or Gemini, forcing companies to duplicate their efforts for each model.

‍

MCP addresses this problem at the root.

It defines a universal and open standard , based on documented conventions shared among vendors.

Developers can thus write once and execute everywhere , without waiting for a particular model to support a new feature.

‍

Furthermore, MCP easily integrates into proprietary or on-premises enterprise environments , allowing organizations to expose their own MCP servers so that models can utilize internal data, without depending on the public cloud .

‍

In Summary:

Proprietary function calling = powerful, but locked into a closed ecosystem.
MCP = open, interoperable, and sovereign, designed for durability.

MCP vs Workflow Automation: From Recipe to Protocol

Workflow automation tools like n8n or Zapier have democratized application integration via pre-defined connectors.

They allow the creation of "if this, then that" scenarios with little to no code, connecting dozens of SaaS tools.

‍

Their limitation: these platforms rely on specific integrations that constantly need updating. They are neither dynamically extensible nor designed to interact directly with language models.

‍

MCP, on the contrary, is an open and extensible protocol , progressively adopted by most LLMs and agent frameworks. It offers developers the ability to describe capabilities once , in a universal format, and execute them everywhere — whether on the cloud or on-premises .

‍

Furthermore, MCP bridges the gap between automation and reasoning : where Zapier executes linear actions, MCP allows an AI agent to plan, choose, and dynamically adapt its actions according to context and feedback received.

‍

In Summary:

Workflow automation = static automation, application-centric.
MCP = dynamic orchestration, centered on intelligent agents and living data .

Use Cases and Adoption

1. Applications Requiring Real-Time Access to Living Data

MCP is ideal for systems where data constantly changes:

BigQuery metrics updated by the minute,
customer orders or inventory in PostgreSQL ,
support tickets in Slack , Zendesk , or Jira ,
or even activity logs and IoT streams .

‍

Thanks to its on-demand invocation capabilities (runtime), MCP allows AI agents to retrieve the most recent information , without pre-processing or re-training, while respecting security and performance constraints.

2. Secure Integration with Sensitive Enterprise Systems

MCP stands out for its native compatibility with modern security standards :

OAuth 2.1 for authentication and delegation,
RBAC (Role-Based Access Control) for permission management,
and ephemeral tokens for granular access control.

‍

This makes it a perfectly suited solution for critical systems — ERP, CRM, financial platforms, private cloud environments — where data must be protected while remaining accessible to AI agents under human supervision.

3. Multi-Step Workflows and Inter-System Orchestration

Complex workflows involving multiple services, databases, or business applications require a protocol capable of coordinating actions in a reliable and traceable manner.

‍

MCP offers this capability through its client-server model with persistent sessions and contextual reasoning capacity .

An agent can thus:

read a resource in a database,
call a tool to transform the data,
then trigger an action in another service,
all within a single transactional cycle .

‍

This is the cornerstone of industrial multi-agent architectures , where multiple intelligences cooperate on distributed tasks.

4. Development of Autonomous Agents with High Reasoning

Agents that must not only execute orders but also reason , anticipate , and adapt according to the context particularly benefit from MCP.

‍

The protocol allows them to combine:

access to structured and verified information sources ,
contextualized business prompts ,
and executable tools to act in the real world.

‍

This combination paves the way for truly autonomous agents , capable of understanding a global business objective and planning their actions across multiple interconnected systems.

5. Interoperability in Multi-Vendor Ecosystems

Finally, MCP addresses a major challenge for large enterprises: avoiding dependency on a single vendor .

Thanks to its open and standardized nature, the protocol allows models from different labs (OpenAI, Anthropic, Google, Mistral, etc.) to interact with the same tools and resources , without recoding integrations each time.

‍

This is a strategic lever for building multi-cloud and sovereign AI environments , where portability and compatibility are guaranteed by the protocol itself.

Pioneers Already Committed

Several technology companies have adopted MCP since its first version, including Block (Square) , Apollo , Zed , Replit , Codeium , and Sourcegraph .

Their common goal: enable their AI agents to deeply understand the application context , interact intelligently with code and systems, and produce more relevant, nuanced, and functional results .

‍

In summary, MCP establishes itself as the universal integration infrastructure for modern AI—a bridge between the logic of models and the complexity of real systems, combining performance, security, and interoperability .

Conclusion

The Model Context Protocol (MCP) represents much more than a technical innovation—it is a major structural advance toward a more mature, composable, and interoperable artificial intelligence infrastructure .

By defining a common language between LLM applications and external systems, MCP transforms a still fragmented landscape into a coherent, governable, and extensible ecosystem .

‍

Thanks to its clear client-server architecture, its structured messages in JSON-RPC 2.0 , its three powerful primitives (resources, prompts, tools), and its flexible transport mechanisms (STDIO, HTTP, SSE), MCP shifts the development of AI applications from a logic of custom and fragile integrations to a modular, standardized, and scalable approach.

Security, Oversight, and Sovereignty

One of the distinctive strengths of the protocol lies in its ethical and operational foundation .

MCP integrates human oversight , explicit consent , and multi-layer security as design principles.

Every interaction between agent and resource is auditable , controlled , and reversible , ensuring that the capabilities it unlocks remain under human governance .

‍

This model allows companies to harness the power of AI agents without sacrificing confidentiality , sovereignty , or traceability —three essential pillars for any responsible adoption of AI in critical environments.

A Rapidly Expanding Ecosystem

The adoption of the protocol is accelerating at an unprecedented pace.

With:

hundreds of official third-party servers , covering major technical and business environments
SDKs available in more than 10 languages (Python, TypeScript, Java, C#, Rust, etc.);
an official registry (MCP Registry) facilitating server discovery and certification;
and the massive support of major industry players — OpenAI, Google DeepMind, Microsoft, AWS ,

‍

MCP is rapidly establishing itself as the de facto standard for context integration in agentic AI systems .

An Infrastructure for the Next Generation of Distributed Intelligences

By making models capable of interacting with live data, business tools, and enterprise systems in a standardized, secure, and governable manner, MCP opens the way to a new generation of AI: agentic, contextualized, and responsible .

‍

Its ambition is not to multiply models, but to connect them intelligently —to create an interoperable agent mesh capable of collaborating, learning, and acting in the service of human decision-making.

In this sense, MCP is to AI what TCP/IP was to the Internet : an invisible but essential infrastructure , allowing intelligences to communicate, cooperate, and evolve within a universal framework.

The host acts as the central orchestrator of the system. It is the container application — for example Claude Desktop, Visual Studio Code, Cursor , or any other AI environment — that manages connections between the various components. Its main functions include:

The host is in a way the conductor contextual networking: it ensures that each interaction is traceable, authorized and secure.

The client therefore acts as a intelligent intermediation layer , ensuring that each server can communicate with the LLM without directly exposing its data or compromising the integrity of the system.

The complete specification of the protocol is written in TypeScript , which serves as the source of truth, and a JSON Schema version is automatically generated to facilitate integration with validation, documentation, or SDK generation tools and client/server implementations in other languages.

Beyond its three fundamental primitives ( Resources , Prompts , and Tools ), the Model Context Protocol (MCP) introduces advanced bidirectional mechanisms that allow servers to become true intelligent agents , capable not only of responding to requests but also of initiating them. These capabilities open the way to a more natural and autonomous interaction between the system's components, while maintaining explicit human control at every critical step.

One of the most frequent confusions around the Model Context Protocol (MCP) comes from its semantic resemblance to the context window of large language models.

However, these two notions relate to very different technical levels : the first concerns the orchestration infrastructure , and the second the internal limits of the model itself.

The success of the Model Context Protocol (MCP) relies on a robust technical ecosystem , designed to accelerate its adoption by the open-source community, software vendors, and large enterprises.

This ecosystem provides both the development tools , the reference implementations , and the industrial integrations necessary to build interoperable and secure agentic applications.

The Model Context Protocol (MCP) does not operate in isolation: it integrates closely with the main agent orchestration frameworks on the market, bringing them standardized interoperability , enhanced security , and unified access to a wide range of tools and resources.

Thanks to these integrations, multi-agent architectures gain in coherence, traceability, and extensibility — three essential conditions for the industrialization of agentic AI.

The Model Context Protocol (MCP) is not merely a technical alternative — it redefines how language models interact with their environment.

To measure its scope, it is essential to compare it to three commonly used approaches in the AI ecosystem: Retrieval-Augmented Generation (RAG) , Proprietary Function Calling , and Workflow Automation Tools .

The Model Context Protocol (MCP) was designed to meet a specific need: to allow artificial intelligence agents to interact with the real digital world in a secure, dynamic, and interoperable manner.

Its advantages are particularly evident in environments where data evolves rapidly, governance is strict, and the complexity of flows exceeds the capabilities of a simple static API.

‍

Here are the most relevant application domains for the protocol:

Why the Model Context Protocol?

Sun, 15 Jun 2025 09:00:00 GMT

The Model Context Protocol (MCP) was born out of a simple yet decisive requirement in the AI ecosystem: to reliably and governably link language models to the real world . Introduced by Anthropic in November 2024 and published as an open standard , MCP proposes a common grammar so that LLMs can understand a goal, converse with tools (APIs, databases, business applications) and then perform traceable actions in an operational environment.

Beyond the technical novelty, MCP addresses a structural problem that was slowing the large-scale adoption of agents in companies: the proliferation of ad-hoc connectors and point-to-point integrations, which are costly to maintain and difficult to audit. By normalizing exchanges (message format, critical metadata, journaling) the protocol reduces the N x M effect between models and tools, accelerates integration and lays the foundations for durable interoperability : a necessary condition for moving from prototype to production.

Concretely, MCP creates an orchestration framework in which each interaction can be explained, verified and, if necessary, supervised by a human . This combination of openness (public standard), portability (independence from a single provider) and native auditability makes it a key building block for deploying multi-agent architectures that are robust, compliant and scalable within organizations.

The isolation problem of LLMs

Despite spectacular advances in reasoning and conversation, large language models remain structurally disconnected from their operational environments. In practice they remain locked behind information silos and legacy systems : they see neither up-to-date business data , nor external APIs , nor execution tools (CRM, ERP, document management, cloud services) that would enable them to act in a useful and measurable way.

This technical claustration has a direct impact on the value created. A model only mobilizes what it learned during its training; it can neither query an up-to-date database , nor read a project folder , nor trigger an action (create a ticket, validate a batch, update a field) within a management system. The result: the answer may be brilliant linguistically but disconnected from the real context and unusable in an operational workflow.

Concretely, this translates into:

Pilots that stall at the demonstration stage because they are not integrated with everyday tools;
Decisions that are not traceable , because the model neither writes nor reads in the systems where evidence is stored;
An inability to orchestrate chained tasks (reading a document repository, coherence check, writing a deliverable, updating a reference).

As long as LLMs remain confined to abstract question-and-answer logic , their potential is mechanically limited. Hence the need for a standardized interaction framework — such as MCP — to reconnect AI to the world of live data and enterprise tools with real traceability, governance and capacity for action.

The M x N combinatorial problem

Before the advent of the Model Context Protocol (MCP) , every connection between an AI system and an external data source had to be developed from scratch . This artisanal approach, inherited from the first generations of AI integrations, led to extreme fragmentation of architectures and a genuine scaling nightmare .

To connect M AI models to N business tools or services required designing M x N distinct connectors , each with its own logic, dependencies, exchange formats and security constraints. This combinatorial explosion was not only complex to manage; it made industrialization practically impossible.

The consequences were many:

Prohibitive development costs. Every integration required several weeks of specialized work, mobilizing rare and expensive skills. As the number of models and tools increased, costs grew exponentially.
Chronic technical fragmentation. Connectors created for OpenAI were incompatible with those for Anthropic, themselves different from those for Google Gemini. Teams found themselves duplicating data schemas, pipelines and business logic, leading to a loss of efficiency and increased complexity.
Massive technical debt. Maintaining a network of ad-hoc integrations quickly became a burden: every upstream API update triggered a cascade of fixes in connected systems, slowing deployment cycles and undermining overall stability.
Confusion about roles and responsibilities. AI teams had to simultaneously master the business logic, the peculiarities of the models and the technical specifics of each API. This lack of clear separation of responsibilities diluted their expertise, created organizational bottlenecks and limited the speed of innovation.

In short, before MCP every attempt at AI integration was a sophisticated yet fragile patch-work . Companies built temporary bridges between their models and their tools, without standards or governance, at the cost of growing complexity and almost impossible scalability. MCP was designed precisely to break with this logic , by establishing a common language between models, tools and digital environments .

Standardization inspired by the Language Server Protocol

Faced with this growing complexity, the Model Context Protocol (MCP) introduces a universal architecture . Instead of an explosive interconnection problem of M x N , MCP reduces it to a much more manageable equation: M + N . In other words, each AI model and each business tool no longer need to talk directly to each other — they now communicate through a common language .

The inspiration for this protocol comes from the world of software development, and more specifically from the Language Server Protocol (LSP) created by Microsoft to standardize exchanges between code editors (VS Code, Vim, JetBrains) and syntax analyzers . Before LSP, each editor had to implement specific support for each language — a situation comparable to that of the AI ecosystem before MCP.

MCP applies this same principle of unification to artificial intelligence. Where LSP simplified life for developers, MCP standardizes communication between AI models and external resources (APIs, databases, business tools, cloud services). In this model:

Each tool or data source implements a single MCP server capable of responding to standardized requests;
Each application or AI agent integrates a single MCP client capable of understanding and exploiting these responses;
The whole forms an interoperable ecosystem in which components can be replaced, combined or extended without complete redevelopment.

By unifying exchanges between artificial intelligences and digital systems, MCP abolishes the barriers between proprietary environments. It transforms integration — previously a technical headache — into a standardized, traceable and lasting dialogue , opening the way to a genuine intelligent mesh between models and tools .

Problems solved by MCP

Interoperability

The Model Context Protocol (MCP) is based on a clear philosophy: openness and neutrality . Designed as an open and agnostic protocol , it works with any language model — whether it is Claude, GPT-4, Gemini, Mistral or any other compatible LLM — and can connect to any data source or external tool , without technological dependence.

In other words, MCP does for the AI ecosystem what HTTP did for the Web: it provides a universal dialogue layer between heterogeneous entities. Gone are closed architectures and fragile bridges between competing solutions: a single AI agent can now query a PostgreSQL database, consult a GitHub repository or execute a cloud command , all via a common, documented and interoperable protocol.

Technically, MCP is built on the proven JSON-RPC 2.0 standard, which it enriches with a set of structured metadata — context, source, date, priority, status, justification and confidence level. This approach ensures clear, verifiable and traceable communication between the various actors in the system (agents, tools, databases, services).

Thanks to this foundation, exchanges become not only interoperable but also auditable and durable . MCP does not just connect artificial intelligences to their environments: it gives them a common language that is stable and transparent, capable of crossing technological boundaries and supporting the rise of agentic AI at enterprise scale.

API fragmentation

Before the appearance of the Model Context Protocol (MCP) , developers had to navigate a fragmented and heterogeneous ecosystem .

Depending on use cases, they resorted to OpenAI's function calling , ChatGPT plugins , frameworks like LangChain , or a multitude of custom-developed REST or GraphQL APIs . Each of these approaches imposed its own conventions:

A different schema format for data;
A specific orchestration logic ;
And an independent maintenance cycle , often heavy and fragile.

This technological dispersion led to strong dependency on providers and a lack of overall coherence: integrations were effective locally but incompatible with each other and difficult to industrialize.

MCP puts an end to this complexity by centralizing interactions within a standardized client-server architecture built on simple, universal principles:

MCP servers expose tools, resources and prompts according to standardized schemas. Each server acts as a structured catalog of capabilities accessible to agents.
MCP clients , integrated into models or agents, maintain stateful sessions , allowing them to retain a persistent context from one interaction to the next — a decisive advance for continuous reasoning and long-term supervision.
MCP hosts (such as Claude Desktop , VS Code or Cursor ) orchestrate connections between servers and clients. They apply security policies , manage permissions and ensure that every exchange remains traceable, controlled and compliant.

Thanks to this unified model, MCP not only harmonizes communication between artificial intelligences and their tools: it creates a stable, interoperable and governable infrastructure in which every component — model, application or resource — naturally fits into a coherent ecosystem.

Vendor lock-in

One of the main criticisms leveled at proprietary approaches — such as OpenAI's function calling or ChatGPT plugins — concerned their ecosystem lock-in . A tool or connector designed for GPT-4 was not compatible with Claude or Gemini , forcing companies to make an exclusive provider choice or duplicate their developments for each environment. This structural dependency limited interoperability and slowed the spread of use cases on a large scale.

The Model Context Protocol (MCP) was specifically designed to break with this closed model . As an open standard published under the MIT license , it can be implemented freely by any market player, regardless of the language model or platform. Its objective is clear: to ensure total interoperability and durable portability of tools across different AI ecosystems.

The year 2025 marks a decisive turning point. After its introduction by Anthropic , the protocol was adopted by OpenAI (March 2025) and then by Google DeepMind (April 2025) , confirming the emergence of an industry consensus around a neutral, shared framework. For the first time, the major players in the sector are converging on a common protocol , not controlled by a single actor, capable of serving as the foundation for real interoperability.

However, this momentum of openness remains fragile . As long as Anthropic remains the main maintainer of the protocol without multi-party governance (such as W3C , ISO or an OpenAI Alliance) being established, there is a risk of fragmentation or unilateral control . The durability of MCP will therefore depend on its ability to be structured as a common good , governed collectively, guaranteeing its independence from the commercial strategies of its founders.

Data sovereignty

One of the major strengths of the Model Context Protocol (MCP) lies in its ability to offer companies granular control over access to and flow of data . Where centralized cloud APIs often require transferring information to remote servers — raising issues of compliance, security and sovereignty — MCP introduces a radically different approach: data stays where it is .

Thanks to its flexible architecture, the protocol allows MCP servers to operate both locally (stdio) and through HTTP connections using Server-Sent Events (SSE) . This flexibility gives organizations the option of choosing their deployment mode according to their security and governance constraints:

Deploy MCP servers within their own infrastructures , whether on-premises or private cloud, to meet data residency requirements , notably those linked to GDPR and industry-specific regulations.
Implement modern authorization mechanisms , compatible with OAuth 2.1 standards and decentralized identifier (DID) authentication , reinforcing control over identity and permissions.
Accurately audit access flows , tracing what data are exposed, to which agents, and under what conditions, thanks to task-based access control (TBAC) policies.

This architecture ensures complete traceability and fine-grained control of confidentiality , without compromising operational efficiency. MCP thus reconciles interoperability and governance , allowing companies to harness the power of AI agents while maintaining control over their sensitive data — an essential condition for deploying agentic systems in regulated or critical environments.

Comparison with earlier approaches

MCP vs OpenAI Function Calling

The function calling introduced by OpenAI in 2023 represented a first attempt to give language models a structured ability to act . The principle was simple: the model could invoke predefined functions by generating JSON objects conforming to a given schema. This innovation kicked off interaction between LLMs and external systems, paving the way for concrete use cases (data retrieval, automation, command execution).

But this approach, while effective at small scale, suffers from structural limitations that hinder its adoption in complex environments:

Proprietary coupling. Function definitions and their metadata are tightly linked to the OpenAI ecosystem, making integrations difficult to reuse in other contexts (Claude, Gemini, Mistral, etc.).
Lack of persistence. Each function call is stateless , i.e. independent of the previous one. The model does not retain the logical thread of a session or the memory of intermediate states.
Static discovery. Functions must be declared in advance before launching the session. It is therefore impossible to add new tools or resources dynamically during execution.
Limited scalability. Orchestration rests on the developer, who must manually manage the sequence of calls, synchronization and context management — a task that quickly becomes unmanageable at scale.

The Model Context Protocol (MCP) overcomes these constraints by proposing a session-based and dynamic architecture :

MCP agents operate within stateful sessions , maintaining a continuous context between exchanges.
The available tools can be discovered dynamically at startup, without prior configuration.
The orchestration is standardized ; the protocol supports parallel invocation of multiple functions and smooth coordination between several agents or services.

In summary, MCP transforms the paradigm initiated by OpenAI's function calling: it is no longer a question of occasionally calling a function, but of orchestrating a fluid, traceable and governed conversation between intelligences and systems , within an interoperable and lasting framework.

MCP vs LangChain

LangChain has established itself as one of the reference frameworks for creating AI agents . It offers a complete toolbox for building complex chains of reasoning, integrating multiple language models and exploiting retrieval-augmented generation (RAG) . Thanks to patterns such as ReAct (Reason + Act), LangChain makes it possible to design agents capable of interacting with their environment iteratively, combining reasoning and execution.

However, this functional richness comes with an additional layer of abstraction , which can make execution heavier and debugging more complex. LangChain acts as a meta-framework on top of LLMs, orchestrating calls and intermediate memories, whereas protocols like MCP or OpenAI's function calling operate more directly at the level of exchanges between model and tool.

Key differences include:

Speed. Function calling — and by extension MCP — operates more quickly than LangChain textual agents, because they remove the intermediate layer of linguistic interpretation and limit successive calls.
Transparency. LangChain advocates an explicit approach to reasoning: every thought step, every agent decision can be followed and analyzed. MCP, by contrast, treats tool invocation as a capsulated transaction — privileging standardization and performance over readability of reasoning.
Flexibility. LangChain excels at designing complex and adaptive workflows involving multiple steps of reflection, verification or content generation. MCP adopts the opposite philosophy: reduce complexity by normalizing exchanges to ensure coherence and portability on a large scale.

In reality, these two approaches are not opposed: they can be highly complementary . It is entirely possible to use a LangChain agent to drive complex reasoning chains while invoking tools via MCP . This combination associates the logical control and modularity of the LangChain framework with the interoperability and standardization of the MCP protocol, offering an ideal balance between agility, governance and performance.

MCP vs REST/GraphQL

The REST and GraphQL protocols have been the pillars of the modern web for more than a decade. Designed for stateless exchanges between clients and servers, they have standardized communication between human applications and digital services. Their effectiveness for classic operations — creating, reading, updating and deleting data (CRUD) — is indisputable. However, these paradigms reach their limits in the face of the needs of AI agents , which require continuous, contextual and governed exchanges.

Architectural comparison — Classic AI Assistant vs Agentic Mesh:

Reactivity: Classic assistant responds to a command — Agentic Mesh acts proactively.
Number of agents: Classic assistant uses 1 — Agentic Mesh uses multiple specialized agents.
Coordination: Classic assistant has none — Agentic Mesh enables inter-agent communication.
Human supervision: Classic assistant requires constant supervision — Agentic Mesh requires only occasional validation.
Traceability: Classic assistant has low traceability — Agentic Mesh provides full traceability via action chains.

While REST and GraphQL aim to facilitate single and predictable exchanges between human applications and remote services, the Model Context Protocol (MCP) adopts a different logic, suited to the agentic ecosystem . MCP does not seek to replace REST or GraphQL; it distinguishes itself by its purpose . Where these protocols orchestrate unitary transactions, MCP manages persistent sessions between AI agents and external resources , capable of maintaining a shared and evolving context.

This approach introduces several breakthroughs:

Dynamic discovery of available tools during the initialization phase, with no need for prior static configuration.
A persistent context that preserves the history of exchanges and decisions, enabling consistent and coherent reasoning.
Native bidirectional events , thanks to the use of Server-Sent Events (SSE) , which allow systems to notify agents in real time — a key element for supervision and multi-agent collaboration.

In short, MCP does not replace existing API paradigms: it completes and extends them . It positions itself as the communication layer of the agentic world , the one that allows artificial intelligences to converse with each other and with their digital environments in a standardized, traceable and evolving framework.

Context of emergence and adoption

Adoption timeline

The adoption of the Model Context Protocol (MCP) has been exceptionally rapid, illustrating the willingness of major players in artificial intelligence to converge on a common standard . In less than a year, MCP has gone from an experimental project to an industry benchmark for interoperability between models and tools.

November 2024 — Launch by Anthropic. Anthropic announces the creation of the Model Context Protocol as an open standard, accompanied by SDKs in TypeScript and Python. The stated goal: to enable language models to reliably converse with external tools and data, whatever the provider.
End 2024 — First industrial partners. Technology companies such as Block (Square), Apollo, Replit, Codeium and Sourcegraph quickly integrate MCP into their platforms. These pioneers demonstrate the protocol's value for development, documentation and software productivity use cases.
March 2025 — Adoption by OpenAI. OpenAI's announcement of the official adoption of MCP marks a strategic turning point: for the first time the two main competitors in the market, Anthropic and OpenAI, align on the same technical base, laying the foundations for an interoperable standard across the sector.
April 2025 — Entry of Google DeepMind. Through Demis Hassabis, Google DeepMind confirms integration of MCP into Gemini and its SDKs, hailing a good protocol that is quickly becoming an open standard for the agentic era of AI. This recognition institutionalizes MCP as a reference infrastructure for coordinating intelligent agents.
April 2025 — Alliance with Microsoft. Microsoft partners with Anthropic to co-develop an official C# SDK, designed to facilitate MCP integration into the .NET environment as well as into Copilot Studio, VS Code and Semantic Kernel. This partnership reinforces the protocol's anchoring in professional development tools.
May 2025 — AWS engagement. Amazon Web Services joins the MCP steering committee and launches Strands Agents, an open source SDK compatible with MCP and other emerging standards. This initiative confirms AWS's desire to participate in an open, multi-cloud interoperability ecosystem.
October 2025 — Launch of the MCP Registry by GitHub. GitHub inaugurates the MCP Registry, a centralized hub allowing developers to discover, install and manage MCP servers. More than 40 official servers feature at launch, offered by Microsoft, GitHub, Dynatrace, Terraform and the open source community.

In less than twelve months, MCP has thus gone from an experimental protocol to an industrial standard . Its adoption trajectory reflects a profound shift in the sector: entry into an open agentic era , where collaboration between AI, tools and infrastructures finally rests on a common language.

Ecosystem and network effects

From February 2025 , the ecosystem surrounding the Model Context Protocol (MCP) began to grow at a blistering pace: over 1,000 MCP servers had already been created by the open source community. Just eight months later, in October 2025 , this number continued to increase exponentially , supported by powerful network effects comparable to those seen during the standardization of the Web or the HTTP protocol.

Two major network effects are at work:

Direct network effect. Every new MCP server — whether it is an integration with Slack, GitHub, Salesforce or an internal database — instantly increases the value of the entire ecosystem. Any compatible MCP client can access these new tools without code modification, creating a cumulative effect of interoperability that accelerates the protocol's spread.
Indirect network effect. The adoption of the protocol by the major industrial players such as Google, Microsoft, OpenAI and AWS acts as a powerful signal of trust. This legitimacy in turn attracts more independent developers, start-ups and companies, enriching the diversity of servers and multiplying use cases.

A new creator economy is taking shape around the protocol. Developers are designing and monetizing premium MCP servers , offering advanced features (enhanced security, analytics, vertical integrations) available through marketplaces , SaaS subscriptions or sponsorship programs .

Specialized platforms, such as Smithery.ai , are gradually emerging as distribution and certification hubs for MCP connectors. They play the role of trusted third parties , ensuring the quality, compatibility and security of servers offered to the community.

In less than a year MCP has thus gone from a technical protocol to an economic engine . Its adoption is no longer driven solely by engineering logic but by an open market dynamic , where interoperability rhymes with shared innovation and collective value creation.

Economic and strategic issues

Reduced integration costs

Economically, the Model Context Protocol (MCP) brings about a structural transformation of the cost model for integrations. Where companies once faced a quadratic problem — with M x N integrations to maintain between AI models and business tools — MCP reduces this complexity to a linear problem : M + N integrations now suffice. Each model and each tool only need to implement a single MCP connector to become compatible with the entire ecosystem.

The direct benefits of this standardization are substantial:

Drastic reduction in development time. Integrations that required several weeks of specialized work can now be completed in a few hours, thanks to unified SDKs and standardized schemas.
Significant reduction in technical debt. By eliminating the multiplication of proprietary connectors, MCP simplifies maintenance and reduces the risk of regressions with each API update. Teams gain in stability and predictability.
More intelligent allocation of resources. Freed from repetitive integration tasks, development teams can focus on what really creates value: the business logic, the quality of models and the user experience.

By reducing redundant efforts and lowering maintenance costs, MCP does not just improve technical productivity: it redefines the economics of integration in AI environments, making it scalable, predictable and sustainable .

Standardization and open ecosystems

The Model Context Protocol (MCP) fully embodies the logic of a digital public good . As an open standard , it generates positive externalities that benefit the entire artificial intelligence ecosystem: pooling development efforts, increased interoperability between tools and the emergence of a common language between agents and infrastructures.

This philosophy stands in contrast to the proprietary strategies adopted by some historical players, notably OpenAI , which keeps part of its technologies under closed license. By promoting a collaborative approach, Anthropic positions itself as an ecosystem architect rather than as a simple model provider — a posture that appeals both to the open-source developer community and to large groups seeking technological sovereignty.

Persistent tensions remain, however:

But this openness is not without from dark zones :

Lack of neutral governance. MCP is not yet administered by an international standards body (like W3C or ISO) but remains under the direct responsibility of Anthropic. This dependency raises the question of the protocol's long-term sustainability and the transparency of technical decisions.
Risk of fragmentation. If Anthropic were to impose certain evolutions unilaterally, other actors could react by launching their own competing protocols — similar to the A2A (Agent-to-Agent Protocol) developed by Google — thereby recreating the initial fragmentation that MCP was precisely intended to eliminate.
Competition between standards. MCP is not alone in the field: it coexists with other initiatives such as the Agent Communication Protocol (ACP), A2A and the Agent Network Protocol (ANP). Each tackles a complementary dimension of agentic interoperability (communication, coordination, governance) but this plurality of standards could ultimately divide adoption efforts.

Implications for businesses

The adoption of MCP is not just a technical evolution — it is a strategic bifurcation . It reshapes the way companies design, deploy and govern their artificial intelligence systems.

Accelerating agentic AI. MCP makes it possible to move from static AI , limited to its training corpus, to dynamic and contextual AI , capable of interacting in real time with business data, internal documents or production systems.
Reduction of vendor lock-in. Thanks to its open architecture, MCP gives companies the freedom to switch models — move from GPT-4 to Claude, or from Claude to Gemini — without having to rewrite all their integrations. It becomes a universal abstraction between models and systems.
Enhanced governance and security. By opening access channels between AI and tools, MCP introduces new security risks: supply-chain attacks, prompt injection or malicious servers. Companies must therefore implement strict governance policies, including internal server registries, whitelists and sandboxing mechanisms to limit the exposure surface.
Interoperability by 2027. According to a forward-looking study by Gartner, by 2027 more than a third of agentic AI implementations will combine several agents with complementary skills, collaborating through protocols such as MCP. This evolution heralds the emergence of a true mesh of interoperable intelligences , capable of permanently transforming the structure of information systems.

Prospects and challenges

Towards multi-actor governance?

As the Model Context Protocol (MCP) becomes the de facto standard, the question of its governance becomes central. To avoid unilateral control by a single actor, several voices — from the open-source community, major cloud vendors and regulators — are calling for the creation of a multi-company consortium to steer the evolution of the protocol.

Such an organism could play a decisive role in the sustainability and legitimacy of the standard, in:

Establishing ethical and security frameworks , to ensure that the protocol remains aligned with principles of transparency, sovereignty and data protection.
Coordinating technical evolutions through community working groups where industry players, researchers and open-source developers would participate in defining extensions and fixes.
Ensuring transparency in the standardization process , like the W3C for the Web, in order to limit the risks of fragmentation and ensure interoperability between implementations.

A governance that is open and distributed would thus be the guarantor of a sustainable ecosystem , in which MCP remains a technological common good , and not the strategic instrument of a single company.

Security and trust challenges

The rapid adoption of MCP within companies nevertheless introduces a series of critical risks that must be anticipated and addressed rigorously.

Insufficient authorization. Early versions of the protocol (before March 2025) lacked native permission management mechanisms. Some servers deployed in production still do not incorporate robust authentication, leaving the door open to unauthorized access.
Expanded attack surface. Each new MCP server added to an environment increases the potential surface area exposed to vulnerabilities. Without centralized control, the risk of compromise or data leakage grows proportionally with the size of the network.
Malicious servers. The ease of deploying an MCP server cuts both ways: it fosters innovation but also allows the introduction of shadow servers lacking security, installed without official validation.
Prompt injection. Malicious prompts can exploit the logic of the protocol to trigger data deletions, bypass access rules or cause leaks of sensitive information.

Mitigation best practices

Systematic verification of digital signatures of servers;
Sandboxing execution environments to isolate agents;
Code review and internal validation before production deployments;
Internal registries of approved servers, with updated whitelists;
Continuous auditing of agent logs and behaviors, to detect any drift or anomaly.

The security of MCP therefore does not depend solely on the protocol itself, but on the operational maturity of those who deploy it. It is by combining an open standard, collective governance and cybersecurity discipline that the ecosystem can earn the lasting trust of companies and institutions.

Conclusion: MCP as the infrastructure of agentic AI

The Model Context Protocol (MCP) is not just a technological evolution — it is a foundational infrastructure for the new era of agentic artificial intelligence . By providing a clear answer to the problems of model isolation , ecosystem fragmentation and combinatorial integration complexity , MCP unlocks the true potential of AI agents: to interact with the real world in a standardized, secure and scalable way.

In less than a year, the protocol has moved from concept to industrial adoption. Its rapid integration by Google, OpenAI, Microsoft and AWS , coupled with the exponential growth of its open-source ecosystem , signals an unprecedented convergence in a historically fragmented sector. This dynamic is not trivial: it marks the birth of a common language for collaboration between artificial intelligences, digital tools and business systems.

But for MCP to become the universal protocol of agentic AI over the long term, several challenges remain:

Truly multi-actor governance , ensuring the neutrality of the standard;
End-to-end security , in the face of rising risks of exploitation and misuse;
And strengthened interoperability with other emerging protocols such as A2A, ACP or ANP, to avoid re-fragmentation of infrastructures.

From this perspective, MCP follows in the footsteps of major standards that have shaped digital history: HTTP/TCP-IP for the Internet, USB-C for hardware or HTML for the Web. It is not a product, but a digital public good — an invisible yet essential layer on which the distributed intelligence of tomorrow will be built.

If governed with transparency, rigor and a spirit of cooperation, the Model Context Protocol could become much more than a tool for integration: the technical foundation for collective innovation , open, sustainable and truly serving humanity.

Exponential growth of the MCP ecosystem

Towards a real MCP economy

To secure their deployments, organizations need to adopt a strict governance hygiene :

The Model Context Protocol (MCP): When and Where it is Essential

Tue, 10 Jun 2025 09:00:00 GMT

The Model Context Protocol (MCP): When and Where it is Essential

The Model Context Protocol (MCP) , launched by Anthropic in November 2024 and already adopted by OpenAI , Google DeepMind , and Microsoft , is establishing itself as the future universal standard for connecting artificial intelligence agents to enterprise systems .

Until now, Large Language Models (LLMs) remained confined to isolated environments, unable to interact directly with the real-world data, tools, and workflows of the company. The MCP lifts this structural lock by providing an open, standardized, and interoperable protocol that fluidly connects the cognitive capabilities of models to the operational infrastructure of organizations.

Adoption Momentum and Growth Indicators

Since its open-source release , the Model Context Protocol (MCP) has seen exponential growth , confirming its status as the new standard for interoperability between AI agents and enterprise systems.

Metrics Confirming the Protocol's Traction

6.7 million weekly downloads of the MCP TypeScript SDK , used in front-end and serverless environments.
9 million weekly downloads of the Python SDK , dominant in back-end integrations, research, and automation workflows.
The official GitHub registry already lists 44 verified MCP servers , covering major integrations: GitHub , Playwright , Notion , Stripe , HashiCorp Terraform , PostgreSQL , and Slack .
At the community level, the ecosystem now boasts over 5,500 active MCP servers and 1,100 dedicated GitHub repositories .

Validation by Tech Giants

Microsoft has integrated MCP natively into Windows 11 and Copilot Studio .
Google has added official protocol support in its Agent Development Kit (ADK) .
OpenAI has included MCP in its Agent SDK , ensuring compatibility between ChatGPT, enterprise Copilots, and third-party infrastructures.

A Historical Parallel: MCP as the New HTTP of AI

In the 1990s, the adoption of HTTP by Netscape, Microsoft, and major access providers marked the birth of the modern Web. Today, the Model Context Protocol follows the same trajectory in the field of artificial intelligence: a simple, open, and extensible protocol that connects heterogeneous systems and catalyzes an entire ecosystem around a common grammar.

Pioneering Sectors and Priority Use Cases

Cloud and Software Development

Software development is the most mature adoption ground. Platforms like Replit, Sourcegraph, Zed, and GitHub Copilot have integrated MCP to allow AI agents to interact directly with version control systems (Git), CI/CD tools, and deployment environments. MCP enables agents to generate code adapted to a project's specific architecture, create Git branches, launch automated tests, and autonomously deploy versions.

Health and Life Sciences

In the health sector, MCP transforms clinical decision support by allowing agents to perform CRUDS operations on electronic health records via the FHIR standard. GE HealthCare demonstrated agentic AI concepts based on MCP to assist radiology workflows. Initial studies indicate a 25% reduction in diagnostic errors and a 30% reduction in treatment costs through the use of MCP servers.

Finance and Financial Services

Block (formerly Square) is among the early adopters, having connected its internal financial systems via MCP, reporting significant gains in productivity and decision quality. The protocol allows AI agents to access real-time market data to automatically adjust investment strategies, with projections indicating a 25% reduction in financial losses due to fraud and anomalies.

Industry and Manufacturing

Siemens and General Electric have implemented MCP-based platforms for industrial automation. Johnson & Johnson has deployed an MCP-based predictive maintenance system that reduced downtime by 30% and improved overall equipment effectiveness (OEE) by 25%. The protocol enables agents to monitor equipment performance, adjust conveyor speeds in real-time, and automate defect detection.

Education

EduBase launched one of the first official MCP servers in edtech, allowing educators to dynamically create assessments, plan exams, and analyze results via natural language conversations with Claude. Tamkang University (TKU) developed a community MCP server to automate course monitoring and unify access to fragmented academic systems.

R&D and Scientific Research

MIT and other labs are using MCP to enable AI agents to interact with data management systems, measurement instruments, and simulation platforms — analyzing experimental data, generating new hypotheses, and automatically configuring instruments to perform new experiments.

Adoption Conditions: Model Maturity and Required Capabilities

AI Model Maturity

MCP only achieves its full value when the organization has advanced AI models , capable of leveraging dynamic discovery and multi-tool orchestration . While the first generations of assistants relied on pre-configured workflows chaining static prompts, MCP allows agents to cross a decisive threshold: understanding which tools to use based on the context, planning the execution order of actions, and dynamically adapting to the results obtained. This shift from prompt engineering to agentic reasoning marks the entry into an era of self-orchestrated intelligent systems .

Orchestration Capabilities

To successfully deploy MCP, the underlying architecture must be able to manage stateful sessions , unlike classic REST APIs, which are inherently stateless . The MCP protocol maintains a persistent context between successive actions of an agent. This allows multiple operations belonging to the same task to be logically linked — for example: "Book a flight, then add it to my calendar, and send the confirmation on Slack." Thanks to this persistent session mechanism, the agent retains memory of the context and can resume an interrupted task, correct or readjust its steps, or synchronize multiple tools without logical rupture.

Governance and Security Requirements

MCP servers — true "chokepoints" between AI models and business systems — become critical assets , concentrating access rights to multiple environments. An academic study published in April 2025 highlighted several potential vulnerabilities : malicious code injections in JSON-RPC message flows, compromise of authentication tokens, and insufficient governance of multi-system permissions.

Recommended Security Best Practices:

Robust Authentication : Systematic implementation of OAuth 2.1 with PKCE , regular rotation of API keys, and multi-factor authentication (MFA).
Zero Trust Model : Continuous verification of all communications and strict application of the principle of least privilege .
Granular Access Controls : Hybridization of RBAC (Role-Based Access Control) and ABAC (Attribute-Based Access Control) models for precise contextual permissions.
Full Encryption : Use of standardized cryptography protocols for data in transit and at rest .
Isolation of Sensitive Environments : Use of containerization (Docker, Podman) or lightweight VMs (Firecracker) to limit the effects of a compromise.
Audit and Observability : Centralized logging in SIEM systems, with automated alerts and access traceability.

Example Integration Roadmap for CIO/CTO

Phase 1: Evaluation and Planning (2-4 weeks)

Strategic Audit : Evaluate the current technological architecture, identify high-value, low-risk use cases (report automation, document analysis). Use the 8-critical-constraints decision framework to assess MCP suitability: performance and latency requirements (acceptable if >500ms), security risk tolerance, token economics and cost structure, operational complexity and team capacity, data localization and regulatory compliance, scalability constraints, technical integration complexity, and ecosystem maturity and vendor risk.

Maturity Assessment : Organizations typically require 18-24 months to demonstrate significant competitive advantages, as institutional learning effects accumulate.

Phase 2: Pilot Deployment (4-12 weeks)

Targeted Pilot Projects: Start with low-risk use cases with read-only access to non-critical systems. Establish clear success metrics including operational indicators (completion time, error rate) and strategic indicators (knowledge accumulation, competitive differentiation).

Technical Configuration : Deploy MCP infrastructure (isolated development, staging, production environments), implement OAuth 2.1 with PKCE, configure secret and environment variable management, enable HTTPS with rate limiting, test connection pooling and circuit breakers, establish schema validation and caching.

Compatible Frameworks and Tools : Select from the 12+ available MCP frameworks: OpenAI SDK (native MCP support for agentic applications), LangChain/LangGraph MCP Adapter, Microsoft Semantic Kernel, Google ADK (Agent Development Kit), Vercel AI SDK, CopilotKit, Langflow (open-source visual builder acting as both an MCP client and server).

Phase 3: Scaling Up and Production (3-6 months)

Progressive Deployment : Extend proven patterns to additional use cases and departments, with phased deployment (internal tools → external functionalities → critical applications). Deploy a centralized governance layer acting as a control plane for all MCP server activity: single authentication (issuance of time-limited and scoped credentials), unified governance (access policies defined in one place, uniformly applied), consolidated audit (all tool calls and policy decisions logged in a single system), and tool classification (identify each tool by canonical name, capability tags, data domain, risk tier, and environment scope).

Interoperability and Technical Positioning

MCP vs. REST/OpenAPI

MCP does not replace REST APIs — it adds an AI orchestration layer on top of existing APIs. Key differences: REST/OpenAPI is designed for developers writing code, while MCP is designed for AI agents and LLMs . REST uses manual documentation; MCP enables automatic capability discovery . REST is stateless; MCP maintains stateful sessions . REST is battle-tested over decades; MCP is emergent (November 2024). REST has integrated horizontal scaling; MCP faces session management challenges. REST context is managed manually by developers; MCP has integrated conversational context .

Technical Foundations

MCP relies on JSON-RPC 2.0 for its messages, with three standardized types: requests (bidirectional with ID), responses (same ID as request, result OR error), and notifications (unidirectional without ID for asynchronous updates). The protocol maintains stateful sessions, allowing the client and server to remember previous messages. MCP Capabilities beyond tool calling include: streaming of partial results, OAuth 2.1 authentication, session management, sampling, dynamic tool discovery, structured error handling, and event notifications.

Anticipated Challenges and Limitations

Security and Compliance Challenges

Each MCP server, if misconfigured or granted excessive permissions, can become a critical compromise point , capable of accessing multiple connected systems. Specific challenges include: expanded attack surface (each new server adds a potential gateway to sensitive resources), oversized permissions (a compromised server with extended rights can exfiltrate confidential data), and lack of native SSO support (the current MCP specification does not yet support enterprise authentication protocols like SAML 2.0 or OpenID Connect).

Operational Complexity

The majority of current MCP servers use the STDIO transport , initially designed for local executions. This mode does not meet the requirements of enterprise deployments: single-user authentication (each instance must be launched manually), mandatory co-location (the server and client must reside on the same machine), lack of network policies, and lack of horizontal scalability. For enterprises, the solution involves adopting the HTTP Streamable transport , which allows for remote, scalable, and secure deployment.

Ecosystem Maturity

MCP is still young: less than a year after its launch, its ecosystem remains in the consolidation phase. Companies must evaluate their tolerance for protocol evolution (possible API or schema changes until 2026), anticipate a learning curve for their development teams, and verify the availability of MCP servers adapted to their legacy systems (SAP, Oracle, SharePoint). At this stage, MCP is more suitable for pilot programs or hybrid environments than for massive deployments in critical production.

Token Economics

MCP can generate significant token consumption if its implementation is not optimized. Each server exposes descriptions, metadata, and sometimes voluminous JSON schemas, which are transmitted to the model for contextualization. Recommended optimization strategies: selective caching of tool schemas and metadata, reduction of the number of exposed tools per server (principle of least capability), compression and minimization of prompt descriptions and resources, and active monitoring of token consumption per deliverable or session via internal metrics.

Strategic Recommendations

MCP is essential when:

The organization has multi-step AI workflows requiring coordination between multiple tools and data sources.
Interoperability between AI models is a strategic issue (avoiding vendor lock-in).
The scalability of AI integrations becomes a bottleneck (N×M problem).
Agent autonomy takes precedence over rigid scripted workflows.
Organizational maturity allows absorbing 18-24 months before significant ROI.

Conversely, avoid MCP if: latency requirements are <500ms (high-frequency trading, real-time gaming), guaranteed stability and absolute vendor independence are needed, intolerance to high security risk without the capacity to implement robust governance, or critical legacy systems without available MCP servers.

Evolutionary Perspectives

In less than two years, MCP has crossed adoption thresholds that standards like OpenAPI or GraphQL took five to seven years to reach. This momentum is explained by three drivers: industrial convergence (OpenAI, Anthropic, Microsoft, and Google now use the same integration protocol), universality of agentic logic (every sector seeks to connect agents to dynamic environments), and network effect (the more MCP servers exist, the simpler and more profitable the creation of new agents becomes).

A Transversal Engine for Sectoral Growth

The protocol is not limited to tech: it fuels a profound transformation of data-intensive sectors. The Edge Healthcare AI market is estimated to reach $208.2 billion by 2030 . The global AI Financial Analytics market is expected to reach $11.4 billion by 2027 , with MCP serving as the interoperability foundation between analysis models, ERPs, and regulatory compliance systems.

Towards "AI-native" Architectures

The MCP ecosystem is now guiding system design towards "AI-native" architectures , meaning they are designed for AI agents before human users. In this paradigm: providers will expose their capabilities via standardized MCP servers, clients will delegate their transactions and analyses to connected agents, and inter-company partnerships will be automatically orchestrated according to shared and audited rules. In other words, MCP becomes the economic interface for inter-agent collaboration .

A Risk of Exclusion for Unprepared Organizations

As agents become the new standard for interaction — in supplier relations, customer management, or B2B alliances — organizations lacking these capabilities risk a form of digital isolation . In the near future, not speaking MCP will be like not speaking HTTP at the beginning of the web. Companies that master the protocol will shape the collaborative ecosystems of tomorrow; others will only access them by delegation.

The Model Context Protocol (MCP): When and Where it is Essential

Tue, 10 Jun 2025 09:00:00 GMT

The Model Context Protocol (MCP) , launched by Anthropic in November 2024 and already adopted by OpenAI , Google DeepMind , and Microsoft , is establishing itself as the future universal standard for connecting artificial intelligence agents to enterprise systems .

Behind this rapid adoption lies a silent, yet decisive revolution : MCP redefines how organizations integrate and deploy generative AI at scale .

Until now, Large Language Models (LLMs) remained confined to isolated environments, unable to interact directly with the real-world data, tools, and workflows of the company.

The MCP lifts this structural lock by providing an open, standardized, and interoperable protocol that fluidly connects the cognitive capabilities of models to the operational infrastructure of organizations.

In other words, MCP does not just improve AI: it recomposes its architecture , transforming language models into fully connected agents , capable of acting, reasoning, and collaborating within existing systems.

Adoption Momentum and Growth Indicators

Recent figures attest to massive and rapid adoption, driven by both independent developers and major technology publishers.

Metrics Confirming the Protocol's Traction

6.7 million weekly downloads of the MCP TypeScript SDK , used in front-end and serverless environments.
9 million weekly downloads of the Python SDK , dominant in back-end integrations, research, and automation workflows.
The official GitHub registry already lists 44 verified MCP servers , covering a spectrum of major integrations: GitHub , Playwright , Notion , Stripe , HashiCorp Terraform , PostgreSQL , and Slack .
At the community level, the ecosystem now boasts over 5,500 active MCP servers and 1,100 dedicated GitHub repositories for the protocol.

These indicators reflect a rare phenomenon: simultaneous adoption by open-source developers and major industry players — a momentum comparable to that of the internet's foundational protocols.

Validation by Tech Giants

The integration of MCP by market leaders sends a strong signal regarding its long-term viability :

Microsoft has integrated MCP natively into Windows 11 and Copilot Studio , making the protocol a key component of its AI ecosystem.
Google has added official protocol support in its Agent Development Kit (ADK) , allowing Gemini and associated tools to consume MCP servers directly.
OpenAI has included MCP in its Agent SDK , ensuring compatibility between ChatGPT, enterprise Copilots, and third-party infrastructures.

This convergence among major players creates a structural alignment effect: when the main AI providers adopt the same protocol, it becomes the de facto standard .

A Historical Parallel: MCP as the New HTTP of AI

In the 1990s, the adoption of HTTP by Netscape, Microsoft, and major access providers marked the birth of the modern Web.

Today, the Model Context Protocol follows the same trajectory in the field of artificial intelligence: a simple, open, and extensible protocol that connects heterogeneous systems and catalyzes an entire ecosystem around a common grammar.

This shift is not just technological — it is infrastructural .

MCP is gradually becoming the universal communication layer between agents, models, and applications, laying the groundwork for an Internet of Intelligences where every component, human or machine, can communicate according to a shared language.

Pioneering Sectors and Priority Use Cases

Cloud and Software Development

Health and Life Sciences

In the health sector, MCP transforms clinical decision support by allowing agents to perform CRUDS operations (Create, Read, Update, Search) on electronic health records via the FHIR standard. GE HealthCare demonstrated agentic AI concepts based on MCP in October 2025 to assist radiology workflows — not only to identify anomalies but also to automatically access prior exams or trigger follow-up scheduling. Initial studies indicate a 25% reduction in diagnostic errors and a 30% reduction in treatment costs through the use of MCP servers.

Finance and Financial Services

Industry and Manufacturing

In the manufacturing sector, Siemens and General Electric have implemented MCP-based platforms for industrial automation. Johnson and Johnson has deployed an MCP-based predictive maintenance system that reduced downtime by 30% and improved overall equipment effectiveness (OEE) by 25%. The protocol enables agents to monitor equipment performance, adjust conveyor speeds in real-time, and automate defect detection.

Education

The educational sector is beginning to adopt MCP to transform pedagogical workflows. EduBase launched one of the first official MCP servers in edtech, allowing educators to dynamically create assessments, plan exams, and analyze results via natural language conversations with Claude. Tamkang University (TKU) developed a community MCP server to automate course monitoring and unify access to fragmented academic systems.

R and D and Scientific Research

Adoption Conditions: Model Maturity and Required Capabilities

AI Model Maturity

MCP only achieves its full value when the organization has advanced AI models , capable of leveraging dynamic discovery and multi-tool orchestration .

While the first generations of assistants relied on pre-configured workflows , chaining static prompts , MCP allows agents to cross a decisive threshold:

understanding which tools to use based on the context,
planning the execution order of actions,
and dynamically adapting to the results obtained.

In other words, AI is no longer content with executing an instruction: it reasons about the process , chooses the most relevant strategy, and acts autonomously within a governed framework. This shift from prompt engineering to agentic reasoning marks the entry into an era of self-orchestrated intelligent systems .

Orchestration Capabilities

To successfully deploy MCP, the underlying architecture must be able to manage stateful sessions , unlike classic REST APIs, which are inherently stateless .

The MCP protocol maintains a persistent context between successive actions of an agent. This allows multiple operations belonging to the same task to be logically linked — for example:

Book a flight, then add it to my calendar, and send the confirmation on Slack.

Thanks to this persistent session mechanism, the agent retains memory of the context and can:

resume an interrupted task,
correct or readjust its steps,
or synchronize multiple tools without logical rupture.

This fundamentally conversational and transactional paradigm brings the behavior of MCP agents closer to that of a human collaborator interacting with a complete digital ecosystem.

Governance and Security Requirements

The adoption of MCP in the enterprise is accompanied by increased requirements for security and governance .

MCP servers — true chokepoints between AI models and business systems — become critical assets , concentrating access rights to multiple environments.

An academic study published in April 2025 highlighted several potential vulnerabilities :

malicious code injections in JSON-RPC message flows,
compromise of authentication tokens ,
and insufficient governance of multi-system permissions .

Faced with these risks, companies must adopt a multi-layered security strategy , combining proactive protection, continuous supervision, and reinforced traceability .

Recommended Security Best Practices

Robust Authentication : Systematic implementation of OAuth 2.1 with PKCE , regular rotation of API keys , and multi-factor authentication (MFA) .
Zero Trust Model : Continuous verification of all communications and strict application of the principle of least privilege .
Granular Access Controls : Hybridization of RBAC (Role-Based Access Control) and ABAC (Attribute-Based Access Control) models for precise contextual permissions.
Full Encryption : Use of standardized cryptography protocols for data in transit and at rest .
Isolation of Sensitive Environments : Use of containerization (Docker, Podman) or lightweight VMs (Firecracker) to limit the effects of a compromise.
Audit and Observability : Centralized logging in SIEM (Security Information and Event Management) systems, with automated alerts and access traceability.

These practices, already adopted by pioneering companies, ensure a balance between innovation and compliance . They make MCP not an additional risk, but a catalyst for governable and secure AI .

Example Integration Roadmap for CIO/CTO

Phase 1: Evaluation and Planning (2-4 weeks)

Strategic Audit : Evaluate the current technological architecture, identify high-value, low-risk use cases (report automation, document analysis).

Use the 8-critical-constraints decision framework to assess MCP suitability :

Performance and latency requirements (acceptable if above 500ms)
Security risk tolerance
Token economics and cost structure
Operational complexity and team capacity
Data localization and regulatory compliance
Scalability constraints
Technical integration complexity
Ecosystem maturity and vendor risk

Maturity Assessment : Organizations typically require 18-24 months to demonstrate significant competitive advantages, as institutional learning effects accumulate.

Establishing the Zero Trust Foundation : Explicitly define policies and paths for all MCP interactions, establish supply chain security standards (code signing, SAST).

Phase 2: Pilot Deployment (4-12 weeks)

Targeted Pilot Projects : Start with low-risk use cases with read-only access to non-critical systems. Establish clear success metrics including operational indicators (completion time, error rate) and strategic indicators (knowledge accumulation, competitive differentiation).

Technical Configuration :

Deploy MCP infrastructure (isolated development, staging, production environments)
Implement OAuth 2.1 with PKCE
Configure secret and environment variable management
Enable HTTPS with rate limiting
Test connection pooling and circuit breakers
Establish schema validation and caching

Compatible Frameworks and Tools : Select from the 12+ available MCP frameworks:

OpenAI SDK : Native MCP support for agentic applications
LangChain/LangGraph MCP Adapter : Lightweight wrapper connecting LangChain to MCP toolchains
Microsoft Semantic Kernel : Orchestration SDK integrating AI tools and agents in serverless environments
Google ADK (Agent Development Kit) : Native support for MCP servers
Vercel AI SDK : Connect applications to tools and agents in serverless
CopilotKit : Frontend integration to compliant MCP servers
Langflow : Open-source visual builder acting as both an MCP client and server

Phase 3: Scaling Up and Production (3-6 months)

Progressive Deployment : Extend proven patterns to additional use cases and departments, with phased deployment (internal tools to external functionalities to critical applications).

Production Requirements :

Technical : Operational connection pooling, tested circuit breakers, configured performance monitoring and alerts, full audit logging enabled.
Security : Access controls to virtual servers implemented, active input/output content filtering, egress controls for sensitive data verified, SOC2/HIPAA compliance validated if applicable, security guardrails for dangerous operations tested.
Operational : Tool lifecycle management processes defined, change management for schema updates planned, incident response procedures documented, performance and availability SLAs established, team onboarding and automated provisioning.
Centralized Governance : Deploy a centralized governance layer acting as a control plane for all MCP server activity: Single Authentication (issuance of time-limited and scoped credentials), Unified Governance (access policies defined in one place, uniformly applied), Consolidated Audit (all tool calls and policy decisions logged in a single system), Tool Classification (identify each tool by canonical name, capability tags, data domain, risk tier, and environment scope).
Deployment Architecture : Choose between local servers (stdio), remote servers (SSE/HTTP), or hybrid architecture. Remote MCP servers are the best proxy for enterprise MCP adoption as they require more effort and trust in client demand, typically deployed by large SaaS organizations (Atlassian, Figma, Asana).

Interoperability and Technical Positioning

MCP vs. REST/OpenAPI

MCP does not replace REST APIs — it adds an AI orchestration layer on top of existing APIs. The fundamental differences:

Primary Users : REST/OpenAPI serves developers writing code; MCP serves AI agents and LLMs.
Discovery : REST/OpenAPI requires manual documentation; MCP enables automatic capability discovery.
State Management : REST/OpenAPI is stateless; MCP maintains stateful sessions.
Protocol : REST/OpenAPI uses REST, GraphQL, or gRPC; MCP uses JSON-RPC 2.0.
Maturity : REST/OpenAPI is battle-tested over decades; MCP is emerging (launched Nov. 2024).
Scalability : REST/OpenAPI has built-in horizontal scaling; MCP faces session management challenges.
Context : REST/OpenAPI context is manually managed by the developer; MCP has built-in conversational context.

Technical Foundations

MCP Capabilities : Beyond tool calling (core), the protocol supports streaming of partial results, OAuth 2.1 authentication, session management, sampling, dynamic tool discovery, structured error handling, and event notifications.

Anticipated Challenges and Limitations

Security and Compliance Challenges

One of the major stakes of MCP lies in the security of the servers — true privileged access points between AI agents and the enterprise infrastructure.

Each MCP server, if misconfigured or granted excessive permissions, can become a critical compromise point , capable of accessing multiple connected systems.

Specific challenges include:

Expanded Attack Surface : Each new server adds a potential gateway to sensitive resources.
Oversized Permissions : A compromised server with extended rights can exfiltrate confidential data or execute unauthorized actions.
Lack of Native SSO Support : The current MCP specification does not yet support enterprise authentication protocols (SAML 2.0, OpenID Connect), complicating integration with identity providers such as Okta , Azure AD , or Ping Identity .

To mitigate these limitations, several pioneering companies are already deploying complementary security layers : authentication proxies, strict network isolation, internal server approval registries, and active monitoring via SIEM systems.

Operational Complexity

Operationally, the majority of current MCP servers use the STDIO transport , initially designed for local executions (child processes of a client application).

This mode, while effective for individual development, does not meet the requirements of enterprise deployments :

Single-User Authentication : Each instance must be launched manually, making multi-user scaling difficult to manage.
Mandatory Co-location : The server and client must reside on the same machine, preventing logical or network separation.
Lack of Network Policies : Inability to filter or control flows between agents and servers.
Lack of Horizontal Scalability : STDIO transport supports neither load balancing nor native high availability .

For enterprises, the solution involves adopting the HTTP Streamable transport , which allows for remote, scalable, and secure deployment — but this is still only partially supported by the official MCP SDKs.

Ecosystem Maturity

MCP is still young : less than a year after its launch, its ecosystem remains in the consolidation phase.

Although growth is spectacular, functional maturity and stability of implementations vary by language and use.

Companies must therefore:

evaluate their tolerance for protocol evolution (possible API or schema changes until 2026),
anticipate a learning curve for their development teams,
verify the availability of MCP servers adapted to their legacy systems (SAP, Oracle, SharePoint, etc.).

At this stage, MCP is more suitable for pilot programs or hybrid environments than for massive deployments in critical production.

Token Economics

Finally, MCP can generate significant token consumption if its implementation is not optimized. Each server exposes descriptions, metadata, and sometimes voluminous JSON schemas, which are transmitted to the model for contextualization. Multiplied by dozens of servers and hundreds of calls, these exchanges can inflate operational costs .

Recommended optimization strategies include:

Selective Caching of tool schemas and metadata to avoid reloading them in every session.
Reduction of the number of exposed tools per server (principle of least capability ).
Compression and minimization of prompt descriptions and resources.
Active monitoring of token consumption per deliverable or session via internal metrics.

These measures help maintain the initial promise of MCP — efficient and controlled orchestration — without cost overrun or cognitive overload for the models.

Strategic Recommendations

MCP is essential when:

The organization has multi-step AI workflows requiring coordination between multiple tools and data sources.
Interoperability between AI models is a strategic issue (avoiding vendor lock-in).
The scalability of AI integrations becomes a bottleneck (N x M problem).
Agent autonomy takes precedence over rigid scripted workflows.
Organizational maturity allows absorbing 18-24 months before significant ROI.

Conversely, avoid MCP if:

Latency requirements are below 500ms (high-frequency trading, real-time gaming).
Guaranteed stability and absolute vendor independence are needed.
Intolerance to high security risk without the capacity to implement robust governance.
Critical legacy systems without available MCP servers.

Evolutionary Perspectives

In less than two years, MCP has crossed adoption thresholds that standards like OpenAPI or GraphQL took five to seven years to reach.

This momentum is explained by three drivers:

Industrial Convergence : OpenAI, Anthropic, Microsoft, and Google now use the same integration protocol.
Universality of Agentic Logic : Every sector, from code to healthcare, seeks to connect agents to dynamic environments.
Network Effect : The more MCP servers exist, the simpler and more profitable the creation of new agents becomes.

Market projections indicate that MCP could become the backbone of the AI agent economy , mirroring HTTP for the web or REST for the cloud.

A Transversal Engine for Sectoral Growth

The protocol is not limited to tech: it fuels a profound transformation of data-intensive sectors.

Health and Edge Computing : The Edge Healthcare AI market is estimated to reach 208.2 billion dollars by 2030 . MCP servers play a key role in securely connecting medical devices, diagnostic models, and hospital research databases, both locally and securely.
Finance and Predictive Analytics : The global AI Financial Analytics market is expected to reach 11.4 billion dollars by 2027 , with MCP serving as the interoperability foundation between analysis models, ERPs, and regulatory compliance systems.

These figures do not just reflect an economic opportunity — they illustrate how MCP restructures value chains , by streamlining the passage from data to decision.

Towards AI-native Architectures

The MCP ecosystem is now guiding system design towards AI-native architectures , meaning they are designed for AI agents before human users .

This architectural shift is based on a simple principle: the most frequent interactions will no longer be between humans and applications , but between agents and systems .

In this paradigm:

providers will expose their capabilities via standardized MCP servers,
clients will delegate their transactions and analyses to connected agents,
and inter-company partnerships will be automatically orchestrated, according to shared and audited rules.

In other words, MCP becomes the economic interface for inter-agent collaboration .

A Risk of Exclusion for Unprepared Organizations

They will be less capable of exchanging contextualized data, automating complex processes, or integrating their systems into the highest-performing partner networks.

In the near future, not speaking MCP will be like not speaking HTTP at the beginning of the web. Companies that master the protocol will shape the collaborative ecosystems of tomorrow; others will only access them by delegation.

Agentic AI risks: 5 real dangers to anticipate to stay in control

Tue, 03 Jun 2025 09:00:00 GMT

Since the emergence of generative artificial intelligence , organizations have discovered tools capable of writing, summarizing, translating, and coding at impressive speeds. But a new generation of technologies is on the horizon: Agentic AI . Here, the goal is no longer just to produce content or answer a request, but to delegate an entire mission to an autonomous system — including planning, execution, and adjustment.

Why Agentic AI Is Not Just Another Kind of AI

This technological leap opens up promising prospects: increased productivity, intellectual scalability, and the automation of complex business processes. But like any major disruption, it is accompanied by profound risks. And the more autonomous the agents are, the more side effects can become systemic.

This article offers a detailed analysis of the key risks associated with Agentic AI, the traps to avoid for decision-makers and CIOs, as well as best practices for combining technological power with strategic control.

Understanding What Agentic AI Is, and What It Is Not

Before discussing the risks, it is essential to clarify what Agentic AI truly encompasses. Unlike classic generative AI, which is often limited to responding to a specific, one-off instruction, agentics introduces a new logic: that of the mission rather than the simple task.

Agentic AI is characterized by several fundamental capabilities:

Understanding an objective formulated in natural language , without requiring ultra-precise or coded instructions.
Planning a coherent sequence of actions to achieve this objective, taking into account context and constraints.
Acting autonomously by mobilizing different tools, APIs, or databases, without relying on constant human guidance.
Self-evaluating and adjusting its trajectory, by detecting its own errors or limitations and correcting its behavior.
Collaborating with other agents or with humans, to function as a link in a network rather than as an isolated tool.

The key difference with classic AI assistants therefore lies in the operational mode: where copilots and chatbots remain reactive and guided step-by-step, Agentic AI adopts a proactive approach, focused on a final objective.

But this autonomy, while paving the way for powerful applications, also raises new risks. Without rigorous supervision, an agent can drift, make decisions outside the planned scope, or interact unpredictably with its environment. This is why governance and human oversight must remain at the heart of any agentic strategy.

The Risk of Hallucinations: Invisible Errors with Heavy Consequences

Hallucinations are a well-known phenomenon to language model users: they are false or invented statements , often formulated with great confidence. In the context of Agentic AI, this risk takes on a new and much more worrying dimension.

Indeed, the agent does not limit itself to providing an isolated answer: it constructs a sequence of actions based on its assumptions. If one of these assumptions is wrong from the start, the entire mission can be compromised. A hallucination upstream acts as a foundational error , contaminating the entire downstream process.

Concrete Cases of Hallucinations in a Professional Context

An agent produces a competitive benchmark citing fictitious sources.
It generates a legal report based on non-existent case law.
It makes an operational decision based on misinterpreted data.

These drifts, sometimes difficult to detect immediately, can have heavy consequences: loss of credibility, strategic errors, or even legal liabilities for the organization.

The Most Frequent Causes

Use of an unverified corpus (unreliable web pages, unstructured or obsolete content).
Imprecise or poorly configured instructions , which leave too much room for the model's interpretation.
Absence of intermediate quality control , allowing errors to pass that could have been corrected earlier.

Best Practices for Limiting Hallucinations

Rely on reliable business corpora , derived from validated internal documents, regulatory databases, or recognized scientific publications.
Implement automated verification by control agents or cross-reviews, capable of detecting and signaling inconsistencies.
Maintain systematic human oversight over critical deliverables, to guarantee the reliability of the final outputs.

In summary, hallucinations are not just an anecdotal flaw: they represent a structural failure if they are not anticipated. In an agentic system, the rigor of the sources and the establishment of control loops therefore become essential conditions.

Poorly Framed Autonomy: When the Agent Acts Without Understanding the Context

Autonomy is the promise of Agentic AI. But poorly framed, it becomes a danger . Without a clear framework, an agent can:

Take too many initiatives , beyond what was expected.
Step outside its scope of competence , venturing into decisions that exceed its area of expertise.
Act without business coherence , or even against internal rules and established procedures.

The real danger therefore lies not in autonomy itself, but in the ambiguity of the contract between the human and the agent . When responsibilities, limits, and control mechanisms are not explicitly defined, the agent can drift and compromise the reliability of the system.

Two Forms of Autonomy to Differentiate

Execution Autonomy : the agent acts within a well-defined framework, with precise rules. Its room for maneuver is limited, and its actions are always aligned with objectives set by the human.
Decision Autonomy : the agent is capable of reformulating objectives or prioritizing its actions without direct supervision. It is this level which, if not strictly supervised, can lead to dangerous drifts.

Examples of Observed Drifts

An agent modifies the order of steps in a quality process without consulting anyone, compromising the compliance of the deliverable.
An agent misinterprets an HR rule, generating an error with legal consequences for the organization.

Best Practices for Framing Autonomy

Define a strict functional perimeter, clearly specifying what the agent can do and what remains the exclusive responsibility of the human.
Implement regular supervision with validation of deliverables by a business expert, to maintain constant control over outputs.
Favor "white box" architectures , allowing the agent's reasoning to be audited and its choices to be explained in case of doubt or incident.

Clearly, the autonomy of Agentic AI must not be endured, but governed . Well-framed, it becomes a lever for performance; poorly defined, it transforms into a risk factor.

Hidden Costs: The Illusion of "Free" AI

At first glance, Agentic AI seems irresistible: a quick-to-deploy tool, inexpensive to run, and immediately productive. But in practice, this image quickly fades. Hidden behind the apparent autonomy of agents are indirect costs that many companies underestimate or even ignore.

Types of Costs to Monitor

Initial configuration time
The more generic an agent is, the more effort it requires to be adapted to the business logic, sector-specific exceptions, and formats expected by users. This initial setup, often long and tedious, can delay adoption.

Correction cost
A poorly framed or incomplete deliverable leads to time-consuming back-and-forths: successive adjustments, time wasted on validation, or even complete rejection of the work produced. The illusion of speed quickly disappears if quality is not met from the first iteration.

Loss of team confidence
An AI that is wrong too often, even if it is technically competent, is quickly abandoned by its users. Organizational distrust is then an intangible but heavy cost, as it slows down adoption and reduces return on investment.

Invisible technical cost
Many agents rely on remote models via API calls. If poorly optimized, these flows can generate significant surcharges at scale, especially when query volumes explode.

Best Practices for Limiting These Drifts

Prioritize configurable agents , capable of finely integrating business ontology from the start, to reduce initial friction and increase relevance.
Establish a user feedback system , to refine responses over time and progressively reinforce agent reliability.
Measure costs per deliverable, and not just by the volume of API calls or tokens consumed. The evaluation must focus on the value created (a document, a synthesis note, a recommendation) rather than isolated technical metrics.

In summary, Agentic AI is not free. It requires a thoughtful investment in configuration, supervision, and optimization. This is the price at which it can deliver on its promise of productivity and avoid becoming an invisible drain on the organization.

Loss of Human Control: The Black Box Syndrome

Among the risks associated with Agentic AI, this one is arguably the most insidious: the illusion of control. When the results produced by an agent seem correct, but no one can explain how or why they were obtained, the organization shifts into a fragile and dangerous management mode.

Potential Consequences

Dilution of Responsibility
If an autonomous agent makes a bad decision, who should bear the consequences? The tool that produced the action, the developer who configured it, or the end-user who validated the deliverable? The lack of clear traceability makes governance uncertain.

Impoverishment of Internal Skills
When humans limit themselves to the role of simple validators, they gradually lose their critical analysis and interpretation capacity. Eventually, teams become dependent on the machine and less able to detect errors or challenge results.

Strategic Risk
An AI that always reasons in a standardized way ultimately standardizes thinking. However, a company's competitive advantage often relies on nuance, intuition, and contextual reading — qualities that AI cannot replicate without human intervention.

Best Practices for Maintaining Human Control

Train users to interact intelligently with the AI. The quality of a response often depends on the relevance of the question asked, hence the importance of learning how to formulate and reformulate effective prompts.
Encourage human reformulation of deliverables. The AI can propose a first version, but the human must retain control over the final adjustment, adding their judgment, creativity, and business insight.
Require transparent and auditable agents . Each step of the reasoning must be explainable, documented, and, if necessary, contestable. So-called "white box" architectures offer this guarantee of explainability and strengthen trust in the AI.

Ultimately, the value of Agentic AI lies not only in its automation power but in its ability to strengthen — and not weaken — the cognitive sovereignty of the human.

The Absence of Governance: The Achilles' Heel of Agentics

In discussions surrounding Agentic AI, the focus is often placed on autonomy . But an equally crucial element is sometimes neglected: governance . An Agentic AI can be high-performing, efficient, and fast; yet without a solid governance framework, it becomes a threat — to regulatory compliance, reputation , and the organization's very resilience .

Risks Related to Absent or Weak Governance

GDPR Non-compliance. Poorly processed personal data, lack of consent, or inappropriate storage of sensitive information: all vulnerabilities that expose the company to sanctions.
Auditability Difficulties. Without clear rules, it becomes impossible to trace a decision, understand the origin of an error, or explain a result to a regulator or client.
Legal Exposure. A decision made by an autonomous agent, if not supervised, can directly engage the company's responsibility. The financial and reputational consequences can be severe.

Key Components of Solid AI Governance

A clear supervision policy : who validates what, when, and with what quality criteria?
Integrated explainability mechanisms , allowing the steps of the reasoning to be understood and each choice to be justified.
Continuous monitoring of performance, errors, and behavioral drifts , to adjust the system in real-time and prevent incidents.

Example: DigitalKin's "Agentic Mesh" Approach

DigitalKin proposes an Agentic Mesh architecture where several specialized agents collaborate... and correct each other . Every action is tracked , every deliverable is linked to its sources , guaranteeing complete auditability. And above all, the human retains the final say , in an assumed logic of sovereignty and transparency.

In short, autonomy without governance is a risk , but well-governed autonomy is a strategic opportunity.

Why Transparency Is the Best Safeguard

In the age of Agentic AI, transparency can no longer be considered a simple bonus ; it must become an absolute prerequisite. Without it, no lasting trust can be established between users, regulators, and systems. It conditions not only security and compliance but also adoption and, ultimately, the performance of the agents.

What Transparency Must Cover

The data sources used by the agent, so that every statement can be verified and linked to a reliable corpus.
The reasoning followed to reach a result , with explicit and understandable logic for a business expert.
The limits of the agent, clearly displayed to avoid unrealistic expectations and recall the areas in which the AI is not reliable.
The role of the human in the decision loop , as human supervision remains a pillar of governance and accountability.

In practice, transparency is defined as an equation: intelligibility + traceability + accountability . A well-designed Agentic AI does not need to maintain mystery. On the contrary, it becomes more effective when it is understood , because users can then dialogue with it critically, detect potential errors, and strengthen its value.

By making every step visible, documented, and contestable, transparency proves to be the best safeguard against drifts and the foundation of the trust essential for large-scale adoption.

FAQs — The Risks of Agentic AI: What Every Decision-Maker Must Know

Is Agentic AI riskier than classic Generative AI ?
Yes. Where Generative AI is limited to answering a specific request , Agentic AI is capable of planning and executing actions autonomously. This autonomy increases the potential scope of errors: an isolated hallucination becomes a chain decision, with much heavier consequences.

Can an AI agent be trusted in a regulated context?
Yes, provided that a governable architecture is deployed: systematic human validation, complete traceability of decisions, and integrated supervision mechanisms. Without these safeguards, compliance cannot be guaranteed.

Do you need AI experts to supervise agents?
Not necessarily. It is primarily business experts who play a key role. The challenge is to train them to collaborate with the agent and correctly interpret its outputs, rather than demanding sharp technical skills in AI.

Are hallucinations 100% avoidable?
No. No architecture can completely eliminate them. However, it is possible to drastically reduce them through the use of reliable sources, cross-supervision mechanisms , and human validation of critical deliverables.

How is the true cost of an AI agent calculated?
The cost is not measured solely by the number of API calls or tokens consumed. It must evaluate the value of the final deliverable , the time actually saved (or lost) by the teams, as well as the technical and organizational costs associated with the usage.

Can Agentic AI replace an employee?
No. It can automate repetitive tasks and accelerate certain analysis steps, but it does not replace critical judgment, contextual discernment , or human creativity . Its role is to augment the human, not to supplant them.

Conclusion: Mastering Agentic AI Means Mastering Your Transformation

Agentic AI should not be feared. It must be framed with rigor. Its transformative power is undeniable, but it requires heightened vigilance. To get the best out of it, three conditions are essential:

A rigorous design, anchored in solid business frameworks and focused on usable deliverables.
Strong governance, ensuring supervision, explainability, and shared responsibility.
A culture of human-machine dialogue, where AI is not a substitute but a partner in thought and action.

Agentic AI is not just another technical innovation. It constitutes a profound reconfiguration of how we think, decide, and produce . It redefines the relationship between human and technology, between delegation and sovereignty, between speed and reliability.

And it is precisely because it is promising that it deserves responsible, transparent, and fully controlled integration. The question is therefore not whether Agentic AI will transform organizations, but how we choose to govern it so that it serves our strategic, human, and societal objectives.

Agentic AI: 7 Key Differences from Traditional AI Agents You Need to Know

Sun, 01 Jun 2025 09:00:00 GMT

Why So Much Confusion Between AI Assistants, AI Agents, and Agentic AI?

Artificial intelligence has revolutionized the way we work, interact with systems, and structure information. Yet as these technologies become more widespread, the language surrounding them grows increasingly blurred. Between AI agents, virtual assistants, generative AI, and agentic AI , confusion is common. While all of these concepts belong to the same family, they represent fundamentally different capabilities.

Today, for both users and business experts, it has become essential to clearly distinguish a traditional AI agent from agentic AI . This is not a matter of semantics, but of understanding the ability to delegate an entire mission to a machine , as one would to a digital colleague.

The rise of platforms such as ChatGPT, Copilot, and Claude has popularized the term AI agent , often used to describe a conversational assistant. Yet in most cases, these are AI assistants which, although powerful, do not act autonomously . They are reactive systems that wait for a command before taking action.

By contrast, agentic AI is designed as a proactive system : one that can understand a broader goal, plan the necessary steps to achieve it, act iteratively, and adjust its behavior based on context and results.

AI Assistant: A Reactive and Limited Tool

An AI assistant is a powerful yet limited tool. It relies on large language models (LLMs) to understand an instruction and generate a response. Its value lies in the speed and quality with which it handles specific tasks — but it remains constrained in terms of autonomy.

Key characteristics of an AI assistant

Reactive: it acts only when prompted.
No persistent memory: it cannot retain context beyond a limited session.
Focused on micro-tasks: writing, rephrasing, translating, summarizing.
No planning capability: it cannot organize a sequence of actions on its own.

In short, an AI assistant is a productivity tool, not an autonomous partner .

Defining Agentic AI: Autonomy, Intention, and Planning

Agentic AI represents a true break from traditional assistant models. While the latter simply react to a specific command, agentic AI is built on the ability of a system to reason autonomously — pursuing a defined goal and adjusting its actions over time.

Its core characteristics include:

Understanding a goal expressed in natural language , without requiring overly detailed or coded instructions.
Planning a coherent sequence of actions to achieve that goal, while accounting for context, constraints, and priorities.
Using external tools — APIs, databases, business platforms, or even other agents — to extend its operational reach beyond its internal model.
Self-evaluation mechanisms , allowing it to correct mistakes, adjust its trajectory, and improve performance as it executes.

The Perceive-Plan-Act-Assess (PPAA) Loop

At the heart of agentic AI lies a fundamental dynamic: the PPAA cycle (Perceive-Plan-Act-Assess) . This functional pattern forms the basis of autonomy and enables an agent to manage complex missions end-to-end.

Perceive: The agent analyzes the user's intent and the surrounding context.
Plan: It breaks the goal into actionable steps, organizing the sequence and prioritizing tasks.
Act: The agent mobilizes the necessary tools to execute the planned actions.
Assess: It analyzes the outcomes, measures alignment with the original goal, and adjusts its course if necessary.

This PPAA cycle introduces a form of adaptive autonomy: the agent doesn’t merely execute a command—it learns and refines continuously . That’s what fundamentally distinguishes agentic AI from a simple generative tool or reactive assistant.

Ability to Collaborate with Other Agents

The following 7 dimensions clearly separate a traditional AI agent from agentic AI:

Reactivity: AI Agent responds to a command — Agentic AI anticipates and plans.
Objective: AI Agent is short-term and task-based — Agentic AI is ongoing and goal-oriented.
Autonomy: AI Agent has none — Agentic AI has full autonomy (under supervision).
Context: AI Agent has very limited context — Agentic AI has deep understanding of business logic.
Evaluation: AI Agent has no feedback loop — Agentic AI performs continuous self-assessment.
Collaboration: AI Agent works alone — Agentic AI operates within a network of agents.
Expected Outcome: AI Agent delivers a single response — Agentic AI delivers a complete deliverable.

The Technical Foundations of Agentic AI

Agentic AI is not just an evolution of language models — it is built upon a combination of complementary technological components that provide autonomy, robustness, and seamless integration into business environments.

Language Models (LLMs): They serve as the core understanding engine.
Business Ontologies: They bring structure and rigor, enabling reasoning within a specific domain.
Automated Planning Systems: These systems decompose a goal into logical sequences of actions.
Long-Term Contextual Memory: This enables the agent to retain project history and maintain continuity across missions.
Interconnections with Existing Digital Tools: APIs, databases, applications, and other agents.
Multi-Agent Orchestration Protocols (such as MCP or LangGraph): These coordinate multiple specialized agents.

The Value of Agentic AI for Businesses

Agentic AI doesn't just enhance operational efficiency — it redefines how professionals approach their missions . Its strength lies in the ability to handle complete projects rather than isolated micro-tasks , integrating directly into existing workflows.

Concrete Business Use Cases

Conducting a sectoral literature review : scanning dozens of publications, filtering relevant information, and producing an actionable synthesis.
Running a full competitive benchmark : collecting data from multiple sources, comparing strategies, and highlighting significant gaps.
Drafting a strategic summary note : enriched with documented sources and explicit references, ready for use by a board or executive committee.
Supporting project management: tracking progress on a complex deliverable , triggering follow-ups automatically.

Designing an Agentic AI: The Building Blocks

Agentic AI marks a turning point — it doesn't just respond, it acts, decides, self-corrects , and can even supervise other AIs . For organizations, this represents a profound shift: moving from a tool-based model focused on isolated execution to a digital collaboration that redefines the very nature of work.

This new dynamic is enabled by the Agentic Mesh — a distributed architecture where multiple agents cooperate autonomously, each with its own specialization. Together, they replicate the logic of a human team — handling research, validation, synthesis, and quality control — but at the scale and speed of digital systems .

By integrating such solutions, organizations don't just boost productivity; they rethink how information is processed, how processes are coordinated, and how collective intelligence is mobilized . Agentic AI thus opens the way to a new era — one of augmented enterprises , where humans and machines build performance, creativity, and resilience together.

This distinction is not merely technical; it shapes architectural choices, the scope of delegation, and the role of humans within decision-making loops.

While the term agent has entered mainstream use through modern AI platforms, in the scientific literature (Russell & Norvig, Wooldridge) it refers to an autonomous system capable of perceiving its environment and acting upon it to achieve objectives.

Concrete examples of current use

Let’s look at a few use cases:

A customer service representative relies on a chatbot to answer standard questions.
An HR department uses an assistant to rephrase a job posting.
A marketing project manager requests a summary of an analytical report to save time when preparing presentations.

In all these cases, the same logic applies: the assistant depends entirely on the user to initiate, frame, and validate each step of the process. It remains a situational support tool — useful, but limited — that never reaches true autonomy.

Agentic AI vs. AI Agent: 7 Key Differences

One of the defining strengths of agentic AI lies in its capacity to collaborate with other specialized agents. It doesn’t just execute an isolated task—it can break down a mission into sub-tasks , delegate them to suitable agents, and coordinate their outputs to deliver a coherent whole.

‍

In practice, an agent can:

Split a complex mission into distinct, hierarchical steps.
Assign specific roles to other agents —for instance, one focused on research, another on synthesis, and a third on fact-checking.
Coordinate the entire process , consolidating contributions and ensuring the final result is consistent and complete.

Applied Example

In an R&D context, an agentic AI could receive a complex mission such as “Produce a literature review on gene therapy for osteoarthritis.” It would then:

Identify relevant databases and extract available sources.
Read, classify, and compare the selected publications.
Produce a structured state-of-the-art overview highlighting trends, uncertainties, and emerging directions.
Annotate the document to enhance readability.
Finally, submit the deliverable for human validation —without needing to be prompted at each step.

‍

This illustrates the major difference between a reactive assistant and a cooperative network of agents : agentic AI doesn’t just execute commands—it orchestrates a full process , with a mission-driven and organizationally autonomous logic.

Read this article to learn more about agent networks (Agentic Mesh) .

It is the synergy among these components that gives rise to Agentic AI — a system capable not only of understanding and producing, but also of planning, collaborating, and adapting .

Toward a Transformation of Work

This paradigm shift has profound implications:

Humans no longer need to intervene at every step — only during key moments of validation and supervision.
Day-to-day work moves from manual execution to intelligent monitoring and strategic steering .
The relationship with AI evolves: it is no longer seen as a mere tool, but as a true digital collaborator , capable of working in a network alongside other agents and human teams.

In short, Agentic AI paves the way for a deep reorganization of work — less focused on repetitive tasks, and more oriented toward human value: decision-making, critical analysis, creativity, and strategic insight.

The Human Impact of Agentic AI

Designing an agentic AI is not only a technological challenge — it begins with a deep integration into the organization’s context and business reality . To be truly useful, the agent must understand the specific “rules of the game” of the company, rather than operate as a generic system detached from real-world practice.

Business Ontology and Organizational Context

The first building block is to equip the agent with a clear business ontology — a structured representation of the organization’s knowledge, processes, and language. This includes:

The vocabulary used by teams, with its nuances, acronyms, and industry-specific terms.
The typical goals pursued by each function — whether optimizing a process, ensuring regulatory compliance, or delivering a strategic project.
The constraints and standards governing each mission: quality norms, deadlines, legal compliance, customer expectations, or regulatory requirements.

By embedding this knowledge from the outset, a gentic AI moves beyond data processing to true reasoning within the organization’s own language and logic . This enables it to deliver results that are relevant, actionable, and aligned with corporate strategy.

Conversational Design and Clear Interfaces

An agent’s autonomy does not eliminate the need for a readable and understandable interaction with its users. Good conversational design and an appropriate interface are essential to build trust and facilitate effective human–AI collaboration.

‍

A well-designed agent should therefore:

Explain its choices intelligibly , making visible the reasoning behind its decisions. This explainability helps users follow its logic and identify potential errors.
Provide a clear but professional interface , avoiding oversimplification. The goal is not to hide complexity but to make it usable — users should access key insights without being overwhelmed by unnecessary technical details.
Request human confirmation or validation when uncertainty is high or when a decision carries strategic, legal, or ethical implications. This reinforces the principle that, even when autonomous, the agent remains part of a cooperative and supervised dynamic .

In other words, agent design is not only about technical performance . It must also create a transparent and professional user experience , where AI becomes a readable and governable partner — not an intimidating black box.

Agentic AI: A New Paradigm for Work

Agentic AI is not designed to replace human expertise — it is meant to amplify it . It acts as an extension of reasoning, a digital collaborator with whom humans can engage in dialogue, correction, and strategic guidance.

Rather than substituting human judgment, it creates a cooperative dynamic : humans define objectives, supervise deliverables, and bring nuance, while the agent handles the heaviest cognitive load — gathering, sorting, structuring, and verifying information.

‍

This division of roles makes it possible to:

Respect human expertise , leaving experts in charge of analysis, decision-making, and creativity.
Free up time and attention by reducing the burden of repetitive or technical tasks.
Elevate the human contribution , allowing people to focus more on strategy, creativity, and relationships.

‍

In this sense, Agentic AI doesn’t just transform productivity — it redefines the quality of human work , enabling a more meaningful balance between automation and discernment.

Understanding Agentic AI in 7 Key Questions

What’s the difference between generative AI and agentic AI?
Generative AI is limited to producing content (text, images, code) in response to a specific instruction. Agentic AI , on the other hand, pursues a defined goal through structured reasoning, planned actions, and adaptive execution.
Can you use agentic AI without being a developer?
Yes — as long as the agent is preconfigured for a specific profession and offers a clear, intuitive interface . The challenge is not technical but functional: enabling business experts to use AI directly in their daily work.
Which professions benefit most from agentic AI?
Fields with high information intensity are the most impacted: strategy, research, R&D, HR, legal, and marketing. Wherever large-scale data collection, analysis, and synthesis are critical, agentic AI delivers substantial value.
Is agentic AI reliable?
It becomes reliable when framed within clear business rules, supervised by humans, and governed by strong oversight mechanisms. Reliability depends not only on the technology itself, but on how it is integrated and managed .
To better understand the associated risks, read our analysis of the five major dangers to anticipate .
Can it work with other AI agents?
Yes. Agentic AI can operate within a network of specialized agents, each fulfilling a distinct role (research, synthesis, validation). This collaborative architecture enhances robustness and consistency across outputs.
Does it replace human employees?
No. Agentic AI saves time and expands human capabilities, but it doesn’t replace expertise, nuance, or critical judgment. It should be seen as a digital collaborator , not a substitute for human professionals.
‍ Can agentic AI improve itself?
Yes, in some cases. It can learn from user feedback, adjust priorities, refine selection filters, and propose improved versions of deliverables based on corrections received. This learning capability increases its value over time.

Agentic Mesh: The AI Architecture That Will Revolutionize Collaborative Work

Tue, 27 May 2025 09:00:00 GMT

From AI Assistant to Autonomous Digital Team

Artificial intelligence has already transformed the way we work. With conversational assistants like ChatGPT, Copilot, and Gemini, professionals have discovered the power of generative AI to write, summarize, or automate micro-tasks. Yet, as use cases become more complex, these models reveal their limits: powerful, yes — but isolated, one-off, and linear.

This is where a new architecture emerges — quieter but radically more ambitious : the Agentic Mesh . It's a network of AI agents that work together as a team — communicating, cross-checking, coordinating, and delivering professional-grade results without constant human supervision .

Still relatively unknown to the public, the Agentic Mesh could soon become the technical backbone of AI-assisted work in the years ahead.

What Is an Agentic Mesh? A simple definition and strategic impact

An Agentic Mesh is an architecture in which multiple specialized AI agents cooperate autonomously to achieve a complex goal. It's no longer a single AI responding to a command, but a distributed digital team — coordinated without human intervention.

From Tool-Based AI to Intelligent Network

A traditional AI tool follows a simple pattern: you give an instruction, you get a response . The interaction is linear, with no collective memory or built-in quality control.

The Agentic Mesh completely changes this logic. Here, several specialized agents operate in parallel: one focuses on information extraction, another on verification, a third on synthesis. Each brings its own expertise — but above all, they collaborate .

These agents hand off tasks, evaluate each other's work, and are capable of self-correcting throughout the process . This dynamic creates a true distributed reasoning chain , where every step benefits from multiple layers of review. The result is not just a raw output, but a structured, coherent deliverable — immediately usable in a professional context.

In short, the Mesh behaves not like a single tool, but like a cooperative intelligent network , mirroring a well-organized human team — with role specialization, peer review, and a continuous focus on the quality of the output.

How It Works: Roles, Coordination, and Self-Verification

Within a Mesh, agents don't operate in isolation. Each has a clearly defined specialty and contributes to the overall process much like a member of a human team:

One agent searches for information , scanning relevant sources and extracting key data.
Another verifies reliability , assessing the accuracy of references, data consistency, and potential bias.
A third synthesizes content , structuring information and highlighting key insights.
A fourth adapts the deliverable , reformulating it in the required tone or format — whether a strategic memo, scientific report, or presentation deck.

What makes this system unique is the active cooperation between agents. They continuously exchange information, escalate complex issues to more capable peers, and delegate tasks when they reach their limits. This ability to self-organize introduces an unprecedented level of resilience and efficiency .

The result: greater productivity — but above all, greater coherence, quality, and traceability .

Why the Agentic Mesh Is a Game-Changer for Knowledge Work

The Agentic Mesh is not just a technical breakthrough — it's an organizational response to real-world business challenges. In practice, a valuable mission is never a single task. It requires searching, understanding, verifying, rephrasing, prioritizing, and presenting.

The Mesh reproduces this logic — but at the scale of AI.

Real-World Examples of Complex Missions Executed by a Mesh

The power of an Agentic Mesh isn't measured by its ability to automate simple actions. It shines in the execution of complex, multi-source missions , precisely where traditional AI reaches its limits.

Scientific literature review — Starting from a corpus of 50 academic sources, the Mesh can analyze, filter, and compare relevant publications. Specialized agents cross-check results, verify citations, and produce a structured synthesis that meets scientific standards — while drastically reducing the time required.
R&D tax credit (CIR) documentation — By combining a patent, a state-of-the-art analysis, and the company's internal technical documentation, the Mesh generates a complete, auditable R&D tax credit report . The agents — or Kins — ensure the accuracy of references, regulatory compliance, and clarity of explanations, greatly facilitating dialogue with tax authorities.
Strategic briefing note for an executive board — The Mesh can consolidate internal reports, sector analyses, and market data to produce a decision-oriented strategic memo . The deliverable highlights opportunities, risks, and potential trade-offs — remaining both clear and actionable for executives.

Across these examples, one principle remains constant: the Mesh doesn't just assist experts — it acts as a distributed digital team , capable of handling massive information flows, structuring knowledge, and delivering professional-grade outputs .

Fundamental Differences from a Traditional AI Assistant

Responsiveness: Traditional AI — Responds to a command | Agentic Mesh — Acts proactively
Number of agents: Traditional AI — 1 | Agentic Mesh — Multiple specialized agents
Coordination: Traditional AI — None | Agentic Mesh — Inter-agent communication
Human supervision: Traditional AI — Constant | Agentic Mesh — Occasional, for validation
Traceability: Traditional AI — Low | Agentic Mesh — Complete, through action chains

The DigitalKin Approach: AI That Thinks as a Team

At DigitalKin , the choice is deliberate: we don't rely on generic, interchangeable, and superficial assistants. We design Kins — specialized, autonomous AI agents built to cooperate. Each Kin has a clear mission and a defined role, ensuring higher work quality and results that align precisely with real business needs.

A Multi-Agent Architecture Built Around Deliverables

The Mesh we've designed operates on a collaborative and distributed logic . Several Kins can be mobilized simultaneously toward a single goal, each contributing its unique value:

One Kin drafts the initial content or synthesis.
Another verifies the data , ensuring accuracy and source compliance.
A third reformulates the deliverable in the client's tone, adapting style and vocabulary.
A fourth evaluates clarity, coherence, and business relevance.

This collaborative workflow mirrors how a well-structured human team operates: role specialization, peer review, and iterative refinement. The outcome is a deliverable that is more robust, transparent, and strategically aligned with user expectations.

By embracing this approach, DigitalKin transforms the Mesh into a true distributed digital workshop , where AI doesn't just generate text but creates reliable, contextualized, and actionable value .

Built-In Quality and Control Through the Kin System

Within the Mesh, quality is not a post-process check — it's native to the architecture . The Kin System relies on fractal self-critique loops , allowing each agent to continuously evaluate the relevance of its own output and that of others.

In practice, a Kin can:

Detect inconsistencies — whether in data contradictions, deviations from business rules, or stylistic drift.
Reassign a task to a better-suited Kin , ensuring that every mission is handled by the most competent agent.
Generate an error report , transparently documenting weak points identified and corrections applied.

This approach establishes a distributed, cooperative review process , where every agent becomes not only a producer of value but also an active quality controller.

The result is a more reliable, auditable, and resilient production chain , capable of delivering outputs that meet the strictest professional standards.

Transparency, Traceability, and Business Alignment

One of the Mesh's key strengths lies in its ability to make every production fully understandable and verifiable . Unlike opaque black box AI systems, it embeds mechanisms that ensure clarity and trust at every step.

Every deliverable is audited: The final result isn't a raw output — it's accompanied by validations and checks ensuring compliance with business and regulatory expectations.
Every source is documented: All references used — articles, databases, or internal documents — are explicitly cited and traceable, enabling users to verify the evidence behind any analysis or recommendation.
Every step is intelligible to business experts: Intermediate reasoning, decision criteria, and chosen methods are presented clearly, allowing users to understand, question, and refine the AI's work when needed.

By combining these three dimensions, the Mesh establishes a new standard of transparency , guaranteed traceability , and natural business alignment — creating a foundation of trust between human and machine collaboration.

Agentic Mesh vs. Human Team: Complementarity and Limits

The Agentic Mesh is not designed to replace humans — its purpose is to complement them by handling low-value-added tasks.

Clear Comparison Between a Human Team and an Agentic Mesh

Coordination: Human Team — Slow, prone to friction | Agentic Mesh — Instantaneous, with no delay
Fatigue: Human Team — Yes | Agentic Mesh — No
Creativity: Human Team — High | Agentic Mesh — Low
Traceability: Human Team — Low | Agentic Mesh — Complete
Cost: Human Team — High | Agentic Mesh — Economically scalable
Adaptability: Human Team — Limited by human factors | Agentic Mesh — Constant, multitasking

Tasks to Delegate to the Mesh

The Agentic Mesh excels when handling information-intensive, repetitive activities that require high methodological rigor. Among the missions that can be confidently delegated to its agents are:

Document analysis — scanning vast corpora of scientific publications, technical reports, or industry databases. The Mesh identifies relevant information while filtering out noise and redundancy.
Production of structured technical content — including executive summaries, regulatory reports, or R&D files. Its ability to organize and format information according to strict standards saves considerable time.
Synthesis, benchmarking, and comparative research — by cross-referencing multiple sources to reveal trends, detect discrepancies, and highlight common patterns.
Factual review and verification — through cross-checking mechanisms that minimize the risk of errors and strengthen the reliability of produced deliverables.

In short, the Mesh excels in tasks where speed, accuracy, and standardization create direct business value.

Tasks to Keep Human

Conversely, some dimensions are intrinsically human and cannot — or should not — be delegated to AI. They are what give meaning, direction, and humanity to collective work:

Strategic decision-making , which requires a holistic understanding of context, long-term vision, and the ability to arbitrate beyond measurable facts.
Creativity, design, and storytelling , where originality, sensitivity, and human intuition remain irreplaceable.
Client relationship management , built on listening, empathy, and trust — deeply human qualities that no algorithm can replicate.
Ethical or political arbitration , since some choices involve values, social responsibility, or societal implications that machines cannot resolve alone.

This defines a clear balance: to the Mesh, analytical and procedural tasks ; to humans, meaning, relationship, and creation. It is within this complementarity that the true strength of the augmented enterprise lies.

Current Use Cases of the Mesh in Business

At DigitalKin , the Agentic Mesh is not a futuristic concept — it is already deployed across several strategic sectors, where it is transforming the way research, innovation, and knowledge production are conducted .

Healthcare and Pharmaceuticals

In this field, scientific literature review is often a time-consuming and complex process. The Mesh automates the screening of publications, selecting only those that meet predefined quality and relevance criteria. Beyond filtering, it also performs cross-source consistency checks , comparing results from clinical trials, academic papers, and specialized reviews. The outcome: researchers save considerable time while gaining access to more reliable, better-documented analyses .

Industrial Innovation

In industry, the Agentic Mesh addresses another critical challenge — the exploration and analysis of patents. It can automatically read large patent corpora, compare them to the scientific and technological state of the art, and extract useful insights to guide innovation strategies. This ability is particularly valuable in highly competitive sectors, where the speed and precision of analysis can make all the difference.

The Mesh also streamlines the generation of R&D Tax Credit (CIR) reports . Through a structured and auditable approach, it automates the collection and formatting of required evidence, ensuring both regulatory rigor and productivity gains.

Looking Ahead: Toward an Operating System for Augmented Organizations

The Agentic Mesh does more than enhance existing tools — it lays the groundwork for a new work infrastructure , one that could soon become the true operating system (OS) of the augmented enterprise.

From SaaS to Organic AI

Today's organizations rely on a multitude of software tools — ERP, CRM, BI platforms, and collaborative systems. Each operates in its own silo, with integrations depending on complex, fragile APIs that are difficult to maintain. This fragmentation hinders agility and accumulates technical debt over time.

With the Agentic Mesh, the paradigm shifts:

Agents become the intelligent connectors between systems. Instead of rigid integrations, autonomous entities orchestrate real-time data exchanges.
AI interprets objectives, mobilizes the right tools, acts, verifies, and delivers. Humans no longer need to search for features — the software adapts dynamically to the mission.
Interaction itself evolves: no more clicking through interfaces to execute commands. Users now express missions — prepare a competitive analysis, structure a regulatory report — and the Agentic Mesh delegates the task to a digital team of specialized agents .

This transition marks a profound shift — from command-based software to cooperative software ; from tools that must be manipulated to digital partners capable of working alongside humans. That is the transformation that will make the Agentic Mesh the invisible yet decisive backbone of the augmented enterprise .

FAQs – Understanding the Agentic Mesh and Its Potential

How is the Agentic Mesh different from an AI assistant? An assistant works alone and reacts to commands. The Mesh, on the other hand, brings together multiple agents that cooperate, self-verify, and produce a complete deliverable.
Does the Mesh replace a human team? No. It handles repetitive, analytical, or technical tasks but leaves strategy, creativity, and relationship management to humans.
Is it risky to delegate missions to autonomous agents? Not if governance is clear. A well-designed Mesh includes safeguards such as traceability, cross-validation, and human supervision.
Can an Agentic Mesh already be deployed in business environments? Yes. Architectures like those developed by DigitalKin are already in use across complex industries such as healthcare, finance, and manufacturing.
What are the key advantages? Time savings, improved document consistency, fewer errors, enhanced traceability, and scalable intellectual functions.
Do you need to be an AI expert to benefit from it? No. The Mesh is built to integrate directly into business contexts, with accessible interfaces and immediately usable deliverables.

Conclusion: The Agentic Mesh as the Foundation of Augmented Work

The Agentic Mesh should not be viewed as a mere technical concept. It represents a profound transformation in the relationship between humans, artificial intelligence, and work . It marks the end of isolated AI tools limited to specific, disconnected tasks — and the beginning of cooperative, distributed, and governable AI .

This type of architecture paves the way for a new generation of organizations capable of:

Automating without standardizing , allowing machines to handle operational complexity while preserving the diversity of human approaches and styles.
Gaining speed without sacrificing quality , through interactive loops where algorithmic efficiency combines with human judgment and expertise.
Scaling human thought without losing originality , amplifying analytical and production capacities while preserving creativity, nuance, and the singularity that define human value.

At DigitalKin , this vision is not a distant future — it is an operational reality already embodied in our solutions. The Agentic Mesh sits at the heart of our approach, empowering companies to combine technological strength with collective intelligence — ensuring that machines remain in service of humanity . Tomorrow, this model will become the standard for all organizations seeking to remain competitive while asserting their cognitive sovereignty and social responsibility .

Humans: Essential in Autonomous AI – 7 Strategic Reasons

Tue, 27 May 2025 09:00:00 GMT

Why Humans Remain Decisive in the Age of Autonomy

In the era of autonomous artificial intelligence, the human role remains indispensable . As AI becomes capable of making decisions without immediate supervision, its technical, contextual, and ethical limits make structured human intervention essential. The goal is not to oppose humans and machines, but to design hybrid ecosystems where AI provides speed and consistency , while humans ensure direction, meaning, and accountability .

The Central Role of Humans in Autonomous AI

Despite the spectacular progress of artificial intelligence , humans remain the ultimate control point in autonomous systems. This role is not symbolic — it is structural and essential to ensure that decisions made by machines are safe, relevant, and aligned with human values.

Supervision and Control

Everything begins with defining the objectives .

An AI system, however powerful, has no morality or values — it only optimizes what it is told to. Determining what should be optimized, considering ethics, social values, and long-term consequences, remains an exclusively human responsibility . Without this initial orientation, even a well-designed system can produce catastrophic outcomes.

Next comes real-time intervention , summarized by the concept of Human-in-the-Loop (HITL) . The idea is simple yet powerful: to include a mechanism allowing human intervention at any point in the decision-making loop — whether to validate an action, adjust a parameter, or stop a process. This supervision can be preventive (filtering actions before execution) or corrective (responding quickly to unexpected machine behavior).

Finally, humans provide what AI still struggles to achieve — contextual validation . A machine can analyze massive amounts of data, but it cannot intuitively grasp cultural subtleties, symbols, ambiguities, or the social implications of a decision. Where AI applies rules, humans interpret, nuance, and adapt to the context.

Emerging Best Practices for Supervision

To make supervision genuinely effective, several practices are now emerging:

Decision guardrails: Certain actions must always undergo a mandatory human double-check (red zones), while others — less sensitive — may proceed with post-action auditing (gray zones).
Runbooks and service-level objectives (SLOs): These frameworks define escalation scenarios and specific thresholds for supervision, such as acceptable intervention delays, cancellation rates, or minimum confidence levels for validating a decision.
Explainable logging: Every approval, timestamp, and justification must be transparently recorded. This traceability not only reinforces trust but also provides a solid audit trail for continuous improvement of supervision processes.

Responsibility and Ethics

While AI can automate processes and accelerate decision-making, it still cannot distinguish right from wrong . AI models have no moral compass — they are statistical tools, not conscious entities. It is therefore the human's duty to ensure that the systems they design, supervise, or use act ethically.

This responsibility goes beyond being careful. It means ensuring that every AI decision respects fundamental human principles : human rights, non-discrimination, fairness, and transparency. As ethical guardians, humans intervene where algorithms are blind to moral dilemmas .

But this is also a legal responsibility . When an autonomous system makes a mistake — whether it's a virtual assistant, a self-driving car, or a recruitment algorithm — it is not the machine that is accountable . Accountability lies with its designer, operator, or owner . Humans remain the legal and moral guarantors , a condition essential for maintaining public and institutional trust in AI technologies.

Strengthening Operational Ethics

To ensure responsible AI use, ethics must be embedded in continuous, concrete practices :

Systematic traceability and auditability: Regular ethical reviews, robustness tests, and documentation of sensitive decisions ensure every choice remains explainable and verifiable.
Impact assessment: Organizations can formalize AI Impact Assessments (AIAs) to identify affected stakeholders, analyze potential risks, plan mitigation measures, and define clear recourse mechanisms.
Multidisciplinary ethics committees: Composed of experts in business, data, law, compliance, and security, these committees must have real veto power. Their mission: to ensure that deployed systems align with strategic goals and fundamental ethical principles.

Governance and Lines of Defense

Effective supervision also depends on clear governance and well-defined lines of accountability . An enhanced RACI matrix can specify who owns, operates, monitors, and audits each model — preventing responsibility gaps.

Before delegating fully to AI, it is recommended to use shadow mode — where the system operates in parallel with human teams. This allows comparison, calibration, and correction without operational risk.

Finally, risk mapping linked to each use case can explicitly connect supervision levels to potential impacts. This enables the definition of precise control thresholds , ensuring that the most sensitive decisions always receive human validation .

The Current Limits of Autonomous Systems

The idea of a fully autonomous AI — capable of operating without any human intervention — may sound appealing on paper. But the technical, ethical, and operational reality is quite different. Without human supervision , artificial intelligence systems face major limitations that compromise their effectiveness, safety, and reliability .

A Very Limited Understanding of Context

One of the most fundamental obstacles remains AI's inability to truly understand the context in which it operates. While these systems can analyze data and identify patterns, they possess no awareness, intuition, or common sense . Their reasoning is based on correlation , not on genuine semantic or intentional understanding .

For example, an AI might detect that a message contains an insult based on certain keywords, but it would fail to grasp irony, sarcasm, or cultural nuance . In critical environments such as medicine, law, or defense , this lack of contextual sensitivity can lead to serious errors.

Moreover, AIs suffer from limited adaptability . They can only function within the boundaries defined during their training. When confronted with novel, ambiguous, or out-of-scope situations , they often react inappropriately — or fail to act at all.

Finally, their performance depends entirely on the quality of their training data . If that data is biased, incomplete, outdated, or unrepresentative , the system will produce inaccurate results — and crucially, will never realize it.

Two Additional Angles to Consider

The first concerns goal misalignment . When an AI system optimizes a local metric, it can unintentionally degrade overall performance . This is the danger of perverse incentives or proxy metrics : a system may achieve its numerical target while harming quality, user satisfaction, or even the organization's reputation .

The second relates to the cost of error . The same statistical accuracy does not have the same value across contexts. An error that is tolerable in a marketing scenario could be catastrophic in medicine or law . This asymmetry demands that supervision be proportionate to risk : the more critical the decision, the more graduated and systematic human intervention must be.

Security Vulnerabilities

Autonomous AI systems are also exposed to a wide range of technical and security threats . They can be manipulated, deceived, or hijacked with relative ease if robust safeguards are not implemented.

A common example is data poisoning — the deliberate manipulation of training data to cause a model to behave abnormally or maliciously. Such attacks can be devastating in sensitive fields like cybersecurity, finance, or robotics .

Another critical risk lies in malicious input attacks . Sometimes, a single well-crafted prompt is enough to mislead a model, pushing it to produce false, dangerous, or confidential outputs .

Finally, there is the problem of transparency . Many AI systems — particularly those based on deep learning — operate as black boxes whose decision-making processes are difficult to interpret. This not only complicates auditing but also makes error detection or drift correction nearly impossible without human oversight .

Variable Performance

According to several studies, up to 85% of AI projects fail , often due to poor oversight, unsuitable data, or a lack of human supervision during design or deployment.

This reality underscores a truth that some in the industry prefer to ignore: technological autonomy does not guarantee functional relevance . The raw performance of a model is not a reliable indicator of its accuracy, usefulness, or robustness in complex real-world contexts. Without critical human oversight , error correction , and a clear framework of accountability , AI systems become not only fragile but also potentially dangerous.

Examples of Failures Without Human Supervision

Theoretical limitations of unsupervised AI become tangible when examining real-world cases where the absence of human safeguards led to critical errors or systemic drift.

Microsoft Tay Chatbot (2016): Launched on Twitter to test interactive learning, Tay was quickly hijacked by users who flooded it with toxic messages. Lacking adequate filters, it began reproducing offensive content within hours, forcing Microsoft to shut it down in less than 24 hours.
Tesla Autopilot (2016): In May 2016, a Model S driver was killed after the Autopilot system failed to detect a white truck crossing the road. The investigation revealed insufficient sensor redundancy and inadequate driver vigilance. Since then, the NHTSA has recorded hundreds of incidents involving Autopilot — highlighting the dangers of partially autonomous systems without strict human oversight.
Meta and Algorithmic Moderation: Studies have shown that Facebook's automated moderation systems frequently let harmful or violent content slip through, and in some cases even amplified its spread through recommendation algorithms. These shortcomings underline the need for human control and effective escalation mechanisms.
Amazon's Automated Recruiting System (2018): Amazon discontinued an internal AI recruitment project after discovering it systematically discriminated against female candidates. The bias came from historical training data that reflected gender imbalances. This case illustrated how human supervision and bias audits are essential before any real-world deployment.

The Human-in-the-Loop (HITL) Concept

Definition

The Human-in-the-Loop (HITL) concept refers to a supervised architecture in which humans actively intervene at every critical stage of an AI's decision-making process. The goal is to embed a human checkpoint — to validate, correct, or annotate results — in order to prevent errors , adjust machine behavior , and ensure alignment with business, ethical, and strategic objectives.

Comparison: HITL vs HOOTL vs HATL

HITL (Human-in-the-Loop): High human involvement — systematic and active human control, with human validation at every key stage.
HOOTL (Human-on-the-Loop): Medium human involvement — continuous monitoring, with human intervention only in case of alerts or anomalies. Passive supervision.
HATL (Human-above-the-Loop): Low human involvement — the human acts as a strategic planner, overseeing objectives without being involved in execution. Strategic control.

The 7 Levels of Human Involvement (Sheridan and Verplank)

To define the degree of automation and human intervention, Sheridan and Verplank proposed a 7-level scale :

Full human control — The human performs all tasks without any machine assistance.
Computer-assisted decision — The AI supports the human, who remains the main decision-maker.
Automated suggestion + human validation — The AI proposes; the human confirms before execution.
Autonomous execution with human override — The AI acts; the human can interrupt if necessary.
Autonomous execution with informed supervision — The AI acts; the human monitors without constant intervention.
Near-complete autonomy — The human is informed afterward, with no prior control.
Full autonomy — The AI acts and decides entirely on its own, without human intervention.

This framework makes it possible to calibrate the level of human responsibility according to the context, risk, and required expertise .

Supervision Indicators and Metrics

For human supervision to remain effective, it must rely on precise and measurable indicators . One of the most useful is the override rate , which measures the proportion of AI-initiated actions that are canceled or modified by a human. This directly reflects the system's relevance and reliability .

Another key metric is the escalation rate , indicating how often the AI refers a case for human validation. A rate that's too low may suggest overconfidence , while a high rate indicates insufficient autonomy . The average intervention time is also crucial: it evaluates how quickly an operator can step in to correct or stop a process.

Finally, tracking data drift and model robustness ensures that performance remains stable over time and across new contexts. The quality of AI explanations is another key dimension: systems capable of clearly justifying their decisions foster user trust and broader acceptance.

Best Practices for Effective Human-AI Collaboration

Building an effective collaboration between humans and artificial intelligence relies on a few foundational principles that must be integrated from the design stage.

The first requirement is transparency and explainability . Every decision made by the AI must be understandable to a human and accompanied by an associated confidence level. It's not just about explaining the final output, but also about clarifying the reasoning steps and data sources involved. This level of explainability enables operators to judge whether a recommendation is reliable, contextualized, and actionable , rather than blindly trusting a black box.

Next comes the need for a clear escalation protocol . In case of anomalies or uncertainty, the AI must know when and how to request human intervention . This protocol should be defined in advance: what thresholds trigger an alert, which channels are used to notify operators, and what response times are acceptable. Without this explicit framework, supervision becomes uncertain and accountability blurred.

A third pillar is closed-loop feedback . Human interventions should not remain isolated; they must be reintegrated into the system to enable continuous improvement. Every correction, validation, or adjustment should enrich the AI's knowledge base, strengthen its models, and reduce the likelihood of recurring errors. In this collaborative learning dynamic, the AI gains maturity — and the human reinforces their role as a guide.

Finally, sustainable collaboration is impossible without proper training for human operators . Experts must be trained not only in the technical use of AI but also in supervision, critical thinking, bias detection, and assessment of response validity . AI does not replace human judgment — it depends on it. Strengthening the skills of those overseeing these systems is therefore a strategic investment , ensuring the safety, performance, and ethics of AI deployments.

Frameworks for Structuring Human Supervision

Beyond general best practices, it is essential to rely on solid methodological frameworks to organize human supervision. These frameworks provide clear reference points and prevent governance from depending solely on intuition or common sense.

Hybrid Intelligence Models illustrate this approach. They define precise modes of cooperation between humans and AI, assigning roles according to each one's strengths — computational speed and analytical power for the machine, critical judgment and contextual understanding for the human. By specifying when and how the AI should request human intervention, these models establish a smooth and sustainable partnership .

The EU AI Act adds a crucial legal dimension . It mandates human oversight for all AI systems considered high-risk — such as those deployed in healthcare, justice, or critical infrastructure. This obligation reflects a clear conviction: some decisions cannot be delegated to a machine, however advanced, without explicit human validation or monitoring.

The Ethics Guidelines for Trustworthy AI , published by the European Commission, offer another key reference point. They highlight seven core principles — including robustness, fairness, transparency , and above all, human agency and oversight — to ensure that technology remains aligned with society's fundamental values.

Finally, more specialized models such as PODS or GUMMI have been designed to reinforce supervision in critical environments. They provide regular checkpoints , detailed explainability mechanisms , and built-in safeguards throughout the decision-making process. These frameworks bring an extra layer of discipline and auditability , particularly valuable in sectors where the cost of error is high.

By combining these frameworks, organizations can build robust governance systems that do not simply correct issues after the fact — but instead anticipate and structure decision-making from the outset.

AI-Human Collaboration: Toward a Hybrid Organization

Artificial intelligence should not replace humans but complement them intelligently . That is the promise of hybrid organizational models , where humans and AI collaborate in a fluid, balanced, and productive partnership. This alliance rests on the complementarity of skills — speed, processing power, and consistency for AI; judgment, creativity, empathy, and ethics for humans.

Forward-thinking companies no longer aim to automate humans, but to design hybrid ecosystems where every entity — human or artificial — plays a role suited to its strengths.

Emerging Models

In this model, each task is co-processed by humans and AI, either sequentially or in parallel. The AI proposes; the human adjusts, validates, or enriches.

Example: An AI drafts a report summary; a human expert validates its relevance, refines the tone, and adds professional nuance.

Centaur and Cyborg Models:

Centaur: The human stays in control, and AI acts as a powerful right hand. It's a model based on partial delegation — the AI performs the analyses, while the human makes the decisions. Best suited for managers and analysts.
Cyborg: The AI is deeply integrated into human reasoning. It's a symbiotic duo, where decisions are made in real time through a continuous co-construction between human and AI. Fits highly dynamic or creative environments.

The Centaur model suits managers and analysts best; the Cyborg model fits highly dynamic or creative environments.

In this structure, the AI handles simple or routine cases and escalates ambiguous or critical ones to humans. It's an intelligent hierarchy of supervision .

Commonly used in content moderation, quality control, or cybersecurity , this model saves time while maintaining high rigor.

The Universal Worker approach places an orchestrating AI at the center of a hybrid team. It allocates tasks among specialized AIs and humans, based on expertise, workload, and predefined roles.

This model is close to Agentic Mesh architectures , where intelligence is distributed, and humans act as strategic supervisors or arbiters .

Kolbjornsrud's 6 Principles for Smart AI-Human Collaboration

Based on the research of Professor Lars Kolbjornsrud (BI Norwegian Business School) , these six principles form a compass for organizing effective AI-human collaboration:

Addition — AI should not replace but augment human capabilities, enhancing collective performance.
Relevance — Use AI where it adds the most value: computation, research, data analysis, and automated production.
Substitution — Let AI take over repetitive or tedious tasks, freeing humans to focus on innovation and relationships.
Diversity — Encourage diversity in approaches and user profiles to avoid uniform thinking and strengthen system resilience.
Collaboration — Design processes where humans and AI interact effectively, with clear interfaces and shared responsibilities.
Explanation — AI must provide understandable explanations for its decisions, to foster trust, transparency, and human correction when needed.

Humans as Guardians of Cognitive Sovereignty

As artificial intelligence penetrates every sphere of life — professional, educational, and social — a crucial question arises: do humans still retain their cognitive sovereignty in the age of machines? In other words, are we still the masters of our own thinking, or are we unconsciously delegating our reasoning to algorithms?

In a world where AI generates content, proposes decisions, and influences our choices, humans serve as a last line of defense . We remain the ultimate guarantors of autonomous thought, critical judgment, and intellectual diversity .

FAQs — Humans and Autonomous AI

Why is the human considered indispensable in autonomous AI?

Humans are indispensable because they play several critical roles: they define objectives, supervise systems, validate decisions in complex cases, and assume legal and ethical responsibility for outcomes. Without human involvement, autonomous AI systems lack discernment, morality, and contextual understanding.

What is Human-in-the-Loop (HITL)?

The Human-in-the-Loop model refers to a configuration in which humans intervene actively at every critical stage of an AI system's decision-making process. This means no major action is executed without human validation or adjustment. HITL is essential in sensitive contexts where safety, compliance, or ethics are at stake.

What are the alternatives to HITL?

There are two main alternatives:

HOOTL (Human-on-the-Loop): The human remains in a passive supervisory role — not validating every action, but able to intervene when anomalies are detected.
HATL (Human-above-the-Loop): The human defines strategic directions and operating rules, without directly engaging in operational decisions.

Each model represents an increasing level of AI autonomy , with corresponding implications for control and accountability .

What are the risks of AI without human supervision?

Unsupervised AI can lead to serious consequences, including:

Critical undetected errors , sometimes with significant human or financial impact.
Amplification of biases , reproducing unfair patterns from training data.
Cognitive disengagement of users, who lose analytical capacity and vigilance.
Dilution of responsibility , making it difficult to identify who is accountable in case of an incident.

These risks justify active human involvement , particularly in high-impact systems.

Can we design truly autonomous AIs?

From a technological standpoint, it is already possible to build AI systems capable of operating autonomously in well-defined environments . However, full autonomy without human oversight carries significant risks — loss of supervision, unpredictable behaviors, and decisions misaligned with human values. In most cases, autonomy must be framed by monitoring and validation mechanisms to ensure both safety and accountability.

What is cognitive sovereignty?

Cognitive sovereignty refers to the ability of an individual — or an organization — to maintain independent judgment, critical thinking, and decision-making authority in the face of AI-generated recommendations. It means not blindly following the machine's suggestions, but questioning, adjusting, or rejecting them when necessary. This is a fundamental condition for preserving intellectual freedom in the era of cognitive automation.

Conclusion: Truly autonomous AI but never without humans

Autonomous AI is not self-sufficient. Without humans, it becomes blind to ethics, incapable of reflection, and vulnerable to systemic drift. That's why human supervision must remain embedded at every level .

The future is not about opposition between humans and machines — it lies in a strategic coexistence between algorithmic power and human judgment .

Human-AI Hybrid Model

Tiered Review Systems

Universal Worker (AI Orchestrator)

Agentic AI: The 3 Key Models Deployed by Microsoft, Google, and Meta

Tue, 20 May 2025 09:00:00 GMT

The New Age of Autonomous AI

Artificial intelligence is entering a new phase. After generative AI — capable of producing text, code, or images on demand — agentic AI is emerging as the next frontier . This new generation doesn't just respond — it acts, plans, delegates, and negotiates, becoming an autonomous software actor within the enterprise.

Tech giants Microsoft, Google, a nd Meta have all launched their own offensives in this space. Behind the media hype lie genuine breakthroughs — and limitations that are still poorly understood.

Why Is Agentic AI Generating So Much Buzz?

The arrival of generative AI — through tools like ChatGPT, Midjourney, or Stable Diffusion — triggered a shockwave. For the first time, anyone could create content in seconds from a simple prompt. This usability revolution opened a vast field of experimentation, but its limits quickly became clear.

Generative models, however powerful, run into several obstacles:

Inability to manage long or complex projects , as they excel at short, one-off tasks but struggle to maintain coherence over time.
Lack of contextual consistency and memory, preventing them from fully embedding into business processes or managing project history.
Difficulty executing chained tasks or interacting fluidly with other software systems , which restricts their integration into professional environments.

It is precisely to overcome these limits that agentic AI has emerged. Unlike isolated generative models, it relies on software agents capable of:

Understanding complex business objectives while considering context, constraints, and priorities.
Planning sequences of actions and continuously adjusting them instead of reacting only once.
Self-evaluating and learning from mistakes , improving performance over time.
Collaborating with humans, APIs, databases, and other agents — acting as a full component of a distributed system.

This paradigm shift moves beyond mere assistance toward intelligent, autonomous delegation , where AI becomes an active partner within organizations, able to contribute to real, measurable projects.

From Prompt to Autonomous Loop: The Agentic Revolution

Traditional generative AI functions as a user-centered assistant — it waits for an instruction (the prompt) and returns an answer, without real autonomy or continuity. Agentic AI takes a decisive leap forward. It doesn't just execute a single command — it can manage an entire mission, from understanding the initial need to delivering the final output.

At the heart of this evolution lies a robust operational loop: Perceive, Plan, Act, Evaluate .

This dynamic gives agents unprecedented capabilities:

Handling unexpected situations, by adapting their action plan when faced with missing data, context changes, or new constraints.
Correcting their own mistakes , through mechanisms of self-assessment and continuous feedback.
Coordinating multiple agents , each specialized in a specific task, to handle complex business needs with coherence and efficiency.

This autonomous loop fundamentally transforms the logic of AI: it shifts from a model of simple execution to one of continuous learning and adaptation , where AI becomes an active player in the production cycle. That is the essence of the agentic revolution: the ability to delegate operational responsibilities to AI — while keeping humans at the center of strategic oversight.

The Three Agentic AI Models Shaping 2025

The Personal Copilot Agent

This first model acts as an intelligent extension of the user. Its purpose is simple: to integrate seamlessly into daily tools and assist humans in their routine activities. It operates as a digital copilot, capable of offering contextual suggestions, anticipating certain needs, and accelerating the execution of repetitive or well-defined tasks.

Target: Individual productivity within familiar environments (office tools, Android, web).

Strengths:

Deep integration into existing ecosystems, allowing immediate adoption without major organizational change.
A refined and intuitive user experience, minimizing the learning curve and promoting daily use.
Exceptional efficiency on micro-tasks: drafting emails, generating presentations, conducting quick research, or automating simple workflows.

Limitations:

Limited autonomy: the initiative always comes from the user, who must trigger and frame the task.
Difficulty managing long or nonlinear processes, where continuity and contextual memory are essential.
Lack of governance or supervision mechanisms, which limits use in critical or regulated environments.

Examples: Microsoft Copilot, Google Gemini, Rewind.ai, Rabbit R1

The personal copilot represents a first step toward agentic AI, but it is often seen as an enhanced assistant rather than a truly autonomous or proactive agent. Its strength lies in immediate efficiency, while its weakness remains its inability to manage extended missions or integrate into collaborative workflows.

The Agent Orchestrator (Supervised Multi-Agent Systems)

With this model, we move beyond the linear prompt-response logic into a collective intelligence dynamic. The orchestrator acts as a digital team leader, coordinating multiple agents with distinct skills: one searches for information, another synthesizes it, a third verifies sources and consistency. The human may remain in the loop to steer or validate, but most of the work happens among autonomous entities.

Target: Complex missions requiring coordination between specialized agents.

Strengths:

Modular architecture , where each agent has a defined role, ensuring clarity and high-quality outputs.
Dynamic human-AI collaboration , where the system doesn't just assist but actively contributes to collective production .
Intelligent task distribution, improving scalability and robustness — since one agent's errors can be compensated by others.

Limitations:

Technically demanding setup, requiring fine-tuning and often time-consuming configuration to achieve reliable performance.
Variable results, heavily dependent on model quality, available tools, and the business context.
Limited industrial deployment — most implementations remain experimental, though corporate interest is rapidly growing.

Examples: AutoGen, TaskWeaver, LangGraph, CrewAI, Meta Agents

The multi-agent orchestrator is an exploratory yet highly promising model. It paves the way for collaborative intelligen ce capable of tackling high-value or cognitively demanding missions — though its large-scale adoption will depend on framework maturity and ease of implementation.

The Governable Business Agent

This model represents the most advanced stage of agentic AI: the governable business agent . Designed for critical environments , these agents are tightly aligned with the standards, taxonomies, and practices of a specific domain. They understand specialized vocabulary, integrate regulatory constraints, and reason according to the organization's internal norms.

From their design phase, they embed control, validation, and explainability mechanisms. Every deliverable can be audited, every decision justified, every reasoning step traced. Their added value lies not only in the speed of execution but in the assurance of compliance and trust they bring to business processes.

Target: Industries requiring high levels of reasoning, traceability, and security.

Strengths:

Deep connection to business ontologies , enabling the agent to speak the same language as human experts and adhere to industry standards.
Built-in supervision , with ethical and operational safeguards, native auditability, and governance embedded directly into the system architecture.
Production of professional-grade deliverables , fully aligned with strategic and regulatory expectations, and ready for executive use.

Limitations:

A more demanding initial deployment phase , requiring significant configuration, contextualization, and team training to unlock the agent's full potential.

Example: DigitalKin

The governable business agent represents the most accomplished response to the needs of demanding organizations. It combines scalability, reliability, and compliance , making it an essential asset in sectors where errors are unacceptable and trust is a non-negotiable prerequisite.

Where Do Microsoft, Google, and Meta Really Stand?

Microsoft — Scalable, Integrated Efficiency

Microsoft is pursuing a universal assistant strategy with Copilot, deeply integrated into the Microsoft 365 and Azure ecosystems. The user experience is fluid and contextualized, with the human remaining at the center of action and control.

On the multi-agent architecture side, projects such as AutoGen and related offerings are increasingly explored by clients, with selective deployments in environments where security and compliance are critical.

Bottom line: a productive, well-integrated AI; autonomy and orchestration capabilities are improving, with industrial maturity varying across business contexts.

Google — An Ecosystem of Exploration and Innovation

Google focuses on openness and rapid iteration through Gemini, AI Studio, and Colab. Developers have access to powerful tools for quickly prototyping agents and orchestrators, while Vertex AI provides the foundation for industrialization.

So far, large-scale adoption remains uneven across sectors: strong among GCP-native data/ML workloads, but more gradual in regulated domains where traceability and governance take precedence.

Bottom line: exceptional R&D excellence and fast time-to-prototype; enterprise maturity depends on governance needs and client context.

Meta — Open Source as a Community Engine

Meta plays a leading role through the Llama family and its research publications that structure the open-source ecosystem. Many community frameworks — LangChain, LlamaIndex, and others — orbit around this foundation (without being official Meta projects) and accelerate experimentation with distributed agents and orchestrations.

In terms of industrialization, companies adopt these components selectively, often in hybrid setups (open source + managed services) to meet security and support requirements.

Bottom line: strong collaborative momentum and significant technical influence; ready-to-use enterprise packaging is still maturing for certain use cases.

Enterprise Watchpoints

Despite their technological advances, the models led by Big Tech face several limitations in real-world enterprise deployment:

Business Contextualization: Generic agents require configuration and domain ontologies to align with real business processes.
Traceability and Auditability: Observability and explainability are improving, but requirements vary greatly across regulated industries.
Governance: Human oversight, actionable explainability, and compliance must be designed by default — through defined roles, validations, and audit logs.
Integration Trade-offs: Proprietary platforms offer security, scale, and speed, while open source brings flexibility and portability. The most effective approach is often hybrid. When connecting agents to business tools, the Model Context Protocol (MCP) plays a pivotal role.
From Prototype to Production: Proofs of concept are multiplying, but operationalization demands SLA definition, MLOps/LLMOps practices, strong security, and fine cost management (models, tokens, tool calls).

CIOs and Business Leaders: Key Watchpoints Not to Overlook

As agentic AI gradually finds its place in strategic roadmaps, both business units and CIOs must carefully assess several critical criteria before committing to deployment.

The true level of agent autonomy . It is essential to evaluate not only what the AI can do on its own, but also how effectively it remains under human oversight. Poorly managed autonomy — without clear safeguards — can quickly lead to inefficiencies or unintended outcomes.
Built-in auditability and supervision. This goes beyond a simple activity log. What is required is fine-grained, actionable traceability designed from the start — ensuring that every decision can be reviewed, understood, and corrected if needed.
Deep integration of business ontologies. For an agent to be truly useful, it must reason in the specific language and logic of the business domain, understand its priorities, and respect its constraints. Without that contextual grounding, results risk being superficial or hard to operationalize.
Robust governance, compliance, and security mechanisms. Governance is not just about monitoring — it is about understanding, controlling, and explaining AI decisions. These safeguards must be embedded in the system's architecture, ensuring long-term trust and resilience.

By combining these four dimensions, companies can ensure that the adoption of agentic AI moves beyond a technical experiment toward a sustainable, governed transformation aligned with their strategic goals.

FAQ – What You Need to Know About Agentic AI

Are all copilots true agents?

No. Most current copilots remain reactive assistants — they respond to user requests but lack real autonomy or initiative. They excel at executing isolated tasks but cannot design or follow a path toward a complex goal.

What is the difference between an agent and an agent orchestrator?

An agent acts alone within a defined scope and mission. An orchestrator, by contrast, coordinates several specialized agents — for extraction, verification, synthesis, or rewriting — to achieve a more complex objective. It is the digital equivalent of a project manager distributing and supervising roles across a team.

Is agentic AI compatible with GDPR?

Yes — provided governance is clear and data is handled locally or under strict control. This can include local hosting, pseudonymization mechanisms, or tight management of data flows. Compliance depends as much on architecture as on use cases — a well-designed agent can fully comply with GDPR principles.

Does open source accelerate the rise of agentic AI?

Absolutely. Open source encourages distributed innovation, standard sharing, and transparency. These dynamics fuel experimentation, allow communities to test new architectures quickly, and accelerate adoption across both enterprises and research ecosystems.

Conclusion – The Future of Agentic AI Lies in Usage and Verticalization

The agentic AI revolution is unfolding in the real world — within organizations that learn to extract business value safely, verifiably, and in line with real operational expectations. Microsoft, Google, and Meta are advancing the technical frontier, but the next generation of agentic AI will be shaped by governable, integrated, and domain-specialized architectures — the ones that combine autonomy with trust, and intelligence with accountability.

Model Context Protocol (MCP): The Universal Connector Between AI and External Tools

Tue, 20 May 2025 09:00:00 GMT

As AI becomes an active agent in professional environments, one requirement is becoming essential: connecting it to existing business tools in a reliable, portable, and standardized way . This is precisely what the Model Context Protocol (MCP) offers.

Introduction – Better Connecting AI to the Digital Ecosystem

Artificial intelligences can no longer operate in isolation. In an interconnected digital world where decisions must be fast, contextualized, and traceable , AIs — especially large language models — must be able to access external sources such as business databases, SaaS tools, internal information systems, specialized APIs, or even real-time event streams .

But this connection to the real world introduces new challenges:

How can we ensure reliable data exchange ?
How do we maintain human oversight and auditability ?
How do we avoid confusion between local data, contextual memory, and business truth ?

This is precisely what the Model Context Protocol (MCP) addresses. Introduced by Anthropic in November 2024 , this open protocol establishes the foundation for a new level of interoperability between AI and digital systems . MCP defines a clear, structured, and traceable standard that allows AI models to interact efficiently with complex business environments — without compromising governance or security .

Note: Despite its name, the MCP is not a communication protocol between AI agents . It is an integration language between models and systems — an essential bridge linking AI to real-world data. Agent-to-agent communication, on the other hand, is handled by Google's Agent2Agent (A2A) protocol, designed as a complement to MCP.

What Is the Model Context Protocol (MCP)?

An Integration Protocol Between AI and Business Tools

As AIs evolve into active agents within professional environments, one requirement becomes critical: they must be connected reliably, portably, and through a shared standard to existing business tools. That's exactly what the Model Context Protocol (MCP) delivers.

Instead of developing custom integrations for each application or database, MCP provides a generic communication framework based on JSON-RPC 2.0 , defining how an AI model can interact with external services, databases, or infrastructures.

In other words, MCP acts as a USB-C for AI — a universal connector that simplifies, secures, and standardizes interactions between models (like Claude, GPT, or Mistral) and business tools (like GitHub, Slack, Postgres, or internal APIs).

A Better Governance Framework for AI–System Interactions

The Model Context Protocol (MCP) is more than a standardized data exchange format between artificial intelligences and digital systems — it defines a complete architecture designed to structure and secure all interactions between an AI agent and its environment.

This architecture is built on several key components:

A clear, rigorously documented technical specification , providing development teams with a common framework and preventing inconsistent implementations.
SDKs available in multiple languages (Python, TypeScript, and other popular environments), making it easier for technical communities to adopt and deploy compatible solutions quickly.
A library of server implementations already covering many real-world services — from GitHub and Slack to PostgreSQL — all available as open source. These ready-to-use modules drastically reduce integration complexity and encourage reuse.

Beyond technical efficiency, MCP ensures that every interaction is secure, traceable, and verifiable , while remaining independent of the chosen LLM provider. Whether using GPT, Claude, or Mistral, the protocol provides a layer of stability and portability . This neutrality strengthens technological sovereignty , allowing companies to switch models or vendors without redesigning their entire infrastructure .

Already adopted in demanding industrial environments, MCP reached a major milestone with its integration into Microsoft's Windows AI Foundry platform. This deployment confirms its ambition: to become a universal infrastructure for orchestrating large-scale AI–tool interactions , and a foundational standard for the agentic ecosystem .

Why MCP Is Essential for AI Integration

The rapid rise of AI models within enterprises comes with a major challenge: connecting them efficiently to digital environments . Without a unified framework, every new integration depends on ad hoc development — fragile, costly, and difficult to scale.

Without a standard like the Model Context Protocol (MCP) , integration attempts follow an unsustainable pattern:

Teams must build as many connectors as there are AI x business tool combinations - an NxM matrix that quickly becomes unmanageable. Every new model or tool multiplies dependencies and clutters the architecture.
Data flows often rely on proprietary APIs or custom scripts , which are opaque and difficult to audit. These one-off solutions lack transparency, can't be reused, and generate technical debt that slows innovation.
Human supervision becomes complex : data streams are fragmented, responsibilities unclear, and the absence of a standard limits visibility, control, and governance.

In short, without a common protocol like MCP, AI–system integration turns into a technical maze — expensive to maintain, hard to secure, and risky at scale. MCP addresses this challenge by offering a universal interaction language , designed to simplify, stabilize, and govern the connections between AI and digital ecosystems.

What MCP Fundamentally Changes

The Model Context Protocol doesn't just simplify technical integration — it reshapes how organizations design and deploy their AI ecosystems . With MCP, companies finally gain a structured, standardized, and extensible framework to connect their intelligent systems to existing infrastructures without starting from scratch each time .

This approach unlocks three major strategic benefits:

Simplified Interoperability — MCP acts as a universal bridge: any compatible model can connect to any MCP-enabled tool. Companies are no longer dependent on proprietary orchestration layers or bespoke connectors. This native interoperability reduces complexity and drastically accelerates new AI use-case deployment.
Seamless Business-Level Supervision — With its standardized structure and embedded metadata, MCP makes it possible to track, prioritize, and audit AI-driven actions. Interactions follow clear operational logic understandable by business teams, improving collaboration between technical experts, operational leaders, and governance functions.
AI Architecture Scalability — By making systems modular, portable, and resilient, MCP provides the foundation for scaling AI deployments. Whether managing a complex multi-agent environment or supporting organizational growth, the protocol ensures the stability and robustness of the infrastructure.

In summary , MCP is becoming the backbone of modern AI integration . It guarantees technical resilience while aligning architectures with business strategy. In a world where AI is scaling at unprecedented speed and scope, the Model Context Protocol stands as the essential standard for orchestrating the distributed architectures of tomorrow .

Technical Architecture of the MCP

The Model Context Protocol (MCP) is built on a modular, rigorously defined architecture designed to ensure the fluidity, security, and traceability of exchanges between artificial intelligences and business systems. At the intersection of software engineering and operational requirements, the protocol is structured around three core technical components , each playing a distinct role in the interaction chain.

MCP Server – The Orchestration Core

The MCP Server is the centerpiece of the agentic architecture — the true conductor coordinating, in real time, all interactions between AI agents and the organization's business tools.

Its primary role is to manage the circulation of requests and responses. In practice, it receives a request from a user or an agent, routes it to the correct destination — whether that's another specialized agent or a third-party tool — then collects and redistributes the response in a structured format.

This orchestration ensures that each agent remains focused on its own mission while contributing effectively to the overall outcome.

The MCP Server also manages connections to external systems. It interfaces with APIs, databases, and internal platforms to extend the reach of AI agents. This interoperability is critical: it allows AI to integrate seamlessly into existing enterprise workflows without data silos or redundant processes .

Another key responsibility is traceability . The server maintains comprehensive, timestamped logs of every exchange. These records facilitate audits, enable fine-grained performance analysis, and play a vital role in incident recovery — allowing for fast, secure system restoration when needed.

Finally, the deployment mode of the MCP Server is a strategic choice. It can be:

Self-hosted , offering full sovereignty over data and infrastructure — ideal for organizations subject to strict confidentiality or regulatory constraints.
Cloud-deployed , providing elasticity and scalability to handle fluctuating workloads or rapid growth.
Many enterprises adopt hybrid architectures , combining the security of on-premise hosting with the flexibility of the cloud.

AI Clients – Models That Speak MCP

Next-generation language models (LLMs) no longer operate in isolation. They interact with their environment through structured requests that follow the MCP standard — the common language between AI, business tools, and human agents.

Within this framework, LLMs act as intelligent clients :

They receive clearly defined tasks from the orchestrator — such as data analysis, document drafting, or automated decision-making.
They produce a response — whether text, code, or recommendations — in a standardized, encoded format readable by all other system components. This ensures seamless interoperability and prevents information loss.
They can be connected not only to different model families (GPT, Claude, Mistral, LLaMA, etc.) but also, in some cases, to human agents . This flexibility enables hybrid workflows where AI and human expertise collaborate within the same operational loop.

This abstraction layer is essential: it standardizes interactions between various AIs, regardless of vendor. An organization is therefore no longer locked into a single model — it can combine multiple specialized models based on its needs, or switch providers easily as technology and sovereignty requirements evolve.

In practice, this gives LLMs a kind of universal passport , allowing them to communicate fluidly with each other and with business systems. This standardization is one of the cornerstones of the robustness and long-term viability of modern agentic architectures.

MCP Messages: A Structured and Rich JSON Format

In modern agentic architectures, all exchanges rely on standardized messages . These are encapsulated in an enhanced JSON-RPC format , which serves both as a technical standard and as a governance safeguard .

Each message includes several key layers of information:

Explicit content , describing the requested task, its execution context, and all associated parameters. This clarity eliminates ambiguity in communication and enables agents to collaborate efficiently.
Critical metadata , such as the request's origin, timestamp, priority level, justification for the generated response, and execution status. These contextual details are essential not only to understand what was done, but also why and under what conditions.
A traceable structure , ensuring that every interaction can be archived, reviewed, or audited easily. This capability is especially valuable in industries where regulatory compliance and operational security are non-negotiable.

Adopting this message format goes far beyond mere technical standardization — it provides native auditability , allowing organizations to demonstrate the rigor of their processes and build trust with stakeholders, whether regulators, clients, or partners. In critical sectors such as healthcare, finance, or defense , this combination of transparency and traceability becomes both a regulatory requirement and a competitive advantage .

Quick FAQs About MCP

Even as the Model Context Protocol gains traction across the enterprise AI ecosystem, it is still often misunderstood. Here are answers to the most common questions to clarify its purpose and limits.

Does MCP allow AI agents to cooperate with each other?

No. MCP is not designed to organize direct collaboration between artificial intelligences. Its primary goal is to enable an individual AI agent to access business tools, databases, APIs, or other information systems. It facilitates context access , not agent-to-agent coordination .

For scenarios where multiple AIs must communicate, reason, or coordinate , another protocol is required.

What's the difference between MCP and A2A (Agent-to-Agent)?

The distinction is simple:

MCP (Model Context Protocol) handles interactions between an AI and its digital environment (tools, data, services).
A2A (Agent-to-Agent) , developed by Google , is a complementary protocol specifically designed for communication and cooperation between AI agents.

Put simply: MCP connects one AI to the world , while A2A connects multiple AIs to each other for multi-agent architectures .

Is MCP compatible with different language models?

Yes. Although initiated by Anthropic and optimized for Claude , MCP was designed as an open standard . It is fully interoperable with other LLMs — including GPT, Mistral, or LLaMA — provided these models are wrapped in an MCP-compatible agent .

This makes it a powerful tool for hybrid architectures that combine multiple vendors and technologies.

Conclusion: MCP, the Cornerstone of Integrated AI

As intelligent systems grow increasingly complex, the Model Context Protocol (MCP) stands out as a foundational technical layer . It offers a structured, scalable, and secure answer to a critical challenge: how to connect AIs effectively to their digital ecosystems — business tools, databases, and cloud services — while maintaining security, oversight, and governance .

By unifying exchanges between AIs and existing systems through an open standard , MCP reduces technical complexity, strengthens interaction traceability, and enhances the scalability of industrial AI architectures.

However, it's important to recognize its boundaries . MCP doesn't allow multiple AIs to reason, coordinate, or dynamically delegate tasks to one another. For that, complementary layers are required — notably Google's Agent-to-Agent (A2A) protocol, which enables large-scale inter-agent collaboration .

This broader vision is precisely where DigitalKin's strategy fits in: by going beyond MCP to design a proprietary Agentic Mesh architecture , we're building a framework for intelligent, sovereign, and secure collaboration among specialized AI agents — a distributed, auditable AI deeply aligned with real business priorities.

In short, MCP is a cornerstone of modern AI , but it's only the beginning. The future of integrated and collaborative intelligence will rely on complementary protocols, thoughtful human supervision, and an ongoing commitment to transparency and meaningful impact .

DigitalKin and CCI Lyon Métropole partner to advance agentic AI for research

m.boisis@digitalkin.ai — Tue, 25 Mar 2025 09:29:33 GMT

The Lyon Chamber of Commerce invests in DigitalKin through CCI Capital Croissance and joins forces to accelerate the adoption of transparent, controllable AI in scientific research.

Artificial intelligence is reshaping how businesses compete - 73% of leaders believe AI will strengthen their teams' performance. Yet adoption remains slowed by a lack of internal expertise, technical complexity and ethical concerns.

CCI Lyon Métropole Saint-Étienne Roanne, which represents 170,000 businesses across its territory, has positioned itself as a strategic partner for companies navigating this transformation offering awareness workshops, tailored audits and hands-on support to help organizations deploy AI with clarity and purpose.

As part of this commitment, CCI Capital Croissance, the Chamber's investment arm, has chosen to invest in DigitalKin, a Lyon-based startup redefining the standards of AI for high-stakes environments such as scientific research and healthcare, where reliability and transparency are non-negotiable.

A new generation of AI: controllable, traceable, endorsable

DigitalKin's platform enables organizations to create and configure autonomous AI agents — Kins — capable of executing complex tasks while guaranteeing full traceability, human control and result reliability. Built on a multi-agent architecture that stands in direct opposition to traditional "black box" AI, the technology delivers complete visibility over how AI operates:

Kins are configurable and controllable — professionals define their own methodology, rules, decision thresholds and validation checkpoints. They are traceable and transparent — every action generates detailed reports, allowing professionals to review and justify each step. They are endorsable — the combination of control, traceability and human validation means experts can fully stand behind the results their AI produces.

Proven results in scientific research

In a recent joint communication, Philippe Valentin, President of CCI Lyon Métropole, and Emmanuel Théry, cofounder of DigitalKin, illustrated this approach through a concrete use case: scientific literature review.

Results already observed: a literature review that would take a human researcher months is completed by a Kin in under 20 minutes. Total research exploration time is divided by 10, allowing teams to cover more ground and spend more time advancing their core work. Scientific rigor is maintained without compromise — the Kin guarantees full transparency over its research, analysis and writing process.

"Kins allow researchers to apply their methods with unmatched precision, automatically document every step, and obtain fully defensible results," said Emmanuel Théry. "A researcher can integrate their personal methodology into the system, have our agents execute it, then verify that everything was carried out according to their scientific expectations. The result becomes fully endorsable by the expert."

Beyond research: growing adoption across sectors

DigitalKin's technology has already been adopted by leading organizations across several industries: pharmaceutical research for protocol execution and documentation, R&D tax credit applications, academic research for literature reviews and methodological analyses, and industrial R&D for innovation project management with full decision traceability.

"Artificial intelligence is an unavoidable digital revolution, but every company must approach this turning point with clear-headedness: understand, test, surround yourself with the right partners. That is also how we are adopting it within our own CCI," said Philippe Valentin, President of CCI Lyon Métropole Saint-Étienne Roanne.

About DigitalKin Founded in 2023 in Lyon, DigitalKin is pioneering transparent agentic AI for high-stakes work. A member of the NVIDIA Inception program and backed by leading French innovation partners, the company is currently raising a Series A to accelerate the international deployment of its technology. DigitalKin's mission: create a symbiosis between human expertise and artificial intelligence, where technology amplifies human capabilities while remaining under full, transparent control.

About CCI Lyon Métropole Saint-Étienne Roanne With 100 elected members and 185 mandates across various institutions, CCI Lyon Métropole represents 170,000 businesses. It supports their development and transitions, acts for territorial performance and attractiveness, and advocates for business interests in public decision-making. www.lyon-metropole.cci.fr

Kins: AI agents that integrate expertise and human intelligence

m.boisis@digitalkin.ai — Tue, 25 Feb 2025 09:29:33 GMT

Discover the interview >

dk-corpo

Les Echos: how DigitalKin is democratizing agentic AI for SMEs

DigitalKin Featured in the 2026 Auvergne-Rhône-Alpes French Startup Mapping

Expertise is no longer sold by the hour: DigitalKin reveals its platform at VivaTech 2026

DigitalKin officially listed in the Hub France IA 2026 cartography as a Trusted AI Partner

DigitalKin selected in the France Digitale 2026 mapping of French AI startups

What is the France Digitale AI startup mapping?

Why this selection matters

Agentic AI: the trend the mapping confirms

A French answer to a global need

Explore the mapping and our platform

In 2026, your expertise deserves better

What happens when everyone uses the same AI?

How do our AI agents for R&D and Innovation transform your daily work?

The result? Researchers and engineers are freed from weeks of manual literature search to focus on interpretation and innovation. They talk about us: A committed ecosystem

Toward an exceptional 2026

DigitalKin's CEO on Lyon Décideurs: "Making expertise accessible and abundant"

AI and ethics: inside Lyon's rising agentic AI ecosystem

The Lyon deeptech turning AI into expert AI agents that are reliable, personalized and truly useful

What's Next? Towards an Open Standardization of Distributed Intelligence

Protocol Evolution Perspectives: Modularity and Cross-Vendor Interoperability

Extensions, Plug-ins, and Standard Evolution

1. Multimodality and Streaming: Towards Perceptive Agents

2. MCP Registry: The Backbone of Trust

3. Enhanced Authentication and Authorization

4. Reference Implementations and Multi-Language Interoperability

Open Governance and Improvement Process

MCP as a Building Block for an Agentic Web: Cooperative AIs and Interoperable Mesh

The Emergence of the Agentic Mesh

Interoperability and Competing Protocols

Web of Agents: A Federated Architecture

Regulatory, Ethical, and Technological Sovereignty Challenges

Distributed Governance and Responsibility

Legal Responsibility and Regulatory Compliance

Bias, Fairness, and Transparency

Technological Sovereignty and Strategic Independence

Role of Open Source and Industrial Alliances

Community Dynamics and Industrial Adoption

Predictions and Future Trajectories (2025–2030)

Standardization and Maturity (2025–2026)

Proliferation of Agents and Multi-Agent Orchestration (2026–2028)

Consolidation and Mature Ecosystem (2028–2030)

Economic Value and Industrial Impact

Conclusion: Towards an Open and Governed Agentic Web

The Founding Principles of the Agentic Mesh

1. Google Agent-to-Agent (A2A)

2. Other Emerging Protocols

3. Towards a Multi-Protocol Consensus

Microsoft: The Copilot–MCP Convergence

Hugging Face: Open Standardization of AI Knowledge

Google: Prudent but Strategic Adoption

OpenAI: The Interoperability Turn

Developer Ecosystem: The Living Base of the Standard

Linux Foundation and Multi-Stakeholder Governance

Security, Supply Chain, and Trust Registries

How MCP Transforms AI Architectures?

From Monolithic SaaS to Agentic Mesh: A New Workflow Architecture

The Old Paradigm: Fragmented Integrations and Monoliths

The New Paradigm: Agentic Mesh Orchestrated by MCP

Transformed Workflow: From Human to APIs via AI

Concrete Benefits: Modularity, Auditability, Traceability, Scalability

Modularity: Microservices Architecture for AI

Auditability and Traceability: End-to-End Visibility

Scalability: From Local Experimentation to Cloud Deployment

Security and Governance: Essential Safeguards

1. OAuth 2.1 Authentication: The First Line of Defense

2. RBAC and Granular Permissions: Security Closer to the Business

3. Human-in-the-Loop: Supervision in Critical Workflows

Internal Registry and Governance: Preventing Shadow AI

Comparison with Existing Multi-Agent Architectures

AutoGen (Microsoft): Conversational Collaboration

LangGraph (LangChain): Stateful and Branched Workflows

CrewAI: Teams of Specialized Agents

Comparative Table: MCP vs. Orchestration Frameworks

Advanced Use Cases: From Document Search to Multi-Contextual Automation

Intelligent Document Search

Specialized Business Assistants

Case Study: Goose, Block's (formerly Square) Unified Agent

Multi-Contextual Automation: Integrated Workflow

GitOps and Software Development

The result? Researchers and engineers are freed from weeks of manual literature search to focus on interpretation and innovation.

They talk about us: A committed ecosystem