Une avancée technique discrète mais essentielle : l'adoption des agents en entreprise dépend désormais de la maturité des tuyaux d'infrastructure plutôt que des seules performances brutes des modèles.
La gestion d'état du protocole empêchait le déploiement d'agents IA à grande échelle dans les infrastructures d'entreprise. Passer au stateless supprime le verrou architectural principal pour connecter des LLM aux outils métiers (Salesforce, Slack, bases de données) à travers des clusters distribués.
Jusqu'à présent, le protocole MCP exigeait que les serveurs conservent l'identifiant de session de chaque client entre deux requêtes, posant des problèmes techniques majeurs derrière un répartiteur de charge multi-serveur. La nouvelle version de la spécification adopte un modèle sans état similaire au Web traditionnel, permettant à n'importe quel serveur d'une ferme de traiter la requête d'un agent sans synchronisation complexe.
Impact concret
Les architectes cloud et développeurs backend peuvent désormais concevoir des serveurs MCP sans devoir implémenter de couche de synchronisation de session type Redis entre leurs instances.
Applications concrètes
Connexion à fort trafic d'un agent à une base de données d'entreprise via un cluster Kubernetes distribué ; déploiement d'outils d'automatisation Slack sans gestion de persistance de session au niveau de la passerelle API.
The Model Context Protocol (MCP) is one of the basic building blocks of AI interoperability, giving AI models a secure way to access external data sources and services. It’s the plumbing that lets a chatbot reach into your calendar, your database, or your internal tools, instead of engineers building custom pipes for every connection. Next week, that protocol is getting a significant update, and while it might not be noticeable to end users, it could make a big difference in how the ecosystem develops. The official spec for the new version has been public since May, but we got an unusually clear explanation of the changes Monday morning from the folks at Arcade — a two-year-old startup that’s built its entire business around the work of getting AI agents to actually function inside real companies, letting them securely connect to and act on tools like Gmail, Slack, and Salesforce. Arcade raised $60 million in June on the idea that most AI agents don’t fail because the underlying models are weak but because the infrastructure around them isn’t ready yet, and that’s what this update is trying to address. Essentially, MCP is changing the way it handles session IDs — the little tokens that servers use to remember “ah, this is the same conversation as five seconds ago” — so servers can operate more easily at a larger scale. As Arcade founder Nate Barbettini puts it: [Under the current system] The first time an MCP client like Claude connects to a server, it sends a “hello”: I’m Claude, here’s my version, here are my capabilities. The server replies with its own capabilities and hands back a session ID... From then on, the client sends that session ID on every request so the server knows it’s the same conversation. Sometimes the ID expires, so the client has to notice, request a new one, and carry on.... Picture a real deployment. You’re running a server for millions of users, behind a load balancer whose entire job is to route each request to whatever server in the farm is free, sometimes in a different region. Now every one of those machines has to know about a session ID that some other machine handed out. It’s not impossible, but it’s a serious pain, and it fights the load balancer instead of working with it. In other words, the current setup assumes one server remembers you, but real companies spread traffic across dozens of servers that don’t talk to each other by default, so today’s MCP servers have to do extra work just to keep track of who’s who. That’s been a significant headache for anyone running an MCP server at scale, and part of the reason we haven’t seen more companies ship large-scale, first-party MCP integrations despite all the hype around agentic AI this year. Under the new system, the protocol will take a looser, “stateless” approach to session IDs on the server side, similar to how most ordinary websites already work, which should make the whole system a lot easier to maintain and, in theory, cheaper to run at scale. That’s all pretty technical, but it’s an important reminder that not every part of AI development is moving at breakneck speeds. While model training races ahead, a lot of the technical infrastructure those models need is still subject to the slow log-rolling of standards-body consensus. It really is happening; it’s just a little slower! Topics AI , mcp When you purchase through links in our articles, we may earn a small commission . This doesn’t affect our editorial independence. Russell Brandom AI Editor Russell Brandom has been covering the tech industry since 2012, with a focus on platform policy and emerging technologies. He previously worked at The Verge and Rest of World, and has written for Wired, The Awl and MIT’s Technology Review. He can be reached at russell.brandom@techcrunch.com or on Signal at 412-401-5489. View Bio November 4 Boston Last chance to save up to $190 on TechCrunch Founder Summit. Join 1,000+ founders and VCs at all stages for real-world scaling insights and connections that move the needle. Savings end June 26, 11:59 p.m. PT . REGISTER NOW Most Popular Apple and Google ordered to purge ‘nudify’ apps from App Stores Lucas Ropek Coca-Cola suspended production at its Fairlife dairy after a ransomware attack Zack Whittaker Tesla driver in fatal Texas crash pressed accelerator 100%, NTSB confirms Sean O'Kane Amid hardware legal battle, OpenAI releases a $230 keyboard for Codex Lucas Ropek Anthropic, Blackstone bet the next trillion-dollar AI business is implementation, not just models Rebecca Bellan OpenAI’s first hardware device is reportedly a screenless speaker that can move Lucas Ropek Anthropic’s newest ad is creeping people out Lucas Ropek