MCP's next update aims to make AI agents easier to run at scale

The Model Context Protocol (MCP) is one of the basic building blocks of AI interoperability — the plumbing that lets a chatbot reach into a calendar, a database, or an internal tool instead of engineers wiring up a custom connection for every service. Next week the protocol ships a significant update. It might not be noticeable to end users, but it could shape how the ecosystem develops going forward. The official spec for the new version has been public since May.

On Monday morning, Arcade — a two-year-old startup that connects AI agents to tools like Gmail, Slack, and Salesforce inside real companies — offered an unusually clear explanation of what’s changing. Arcade raised $60 million in June on the premise that most AI agents fail not because the underlying models are weak, but because the infrastructure around them isn’t ready yet. This update, the company argues, targets exactly that gap.

The core change concerns how MCP handles session IDs — the tokens servers use to recognize a returning conversation. Arcade founder Nate Barbettini explained that under the current system, an MCP client like Claude sends a “hello” to a server, which replies with its capabilities and a session ID the client must keep resending on every request; if that ID expires, the client has to request a new one and carry on. At scale, that becomes a serious pain: a server farm behind a load balancer that routes requests across regions has to keep every machine synced on session IDs handed out elsewhere, fighting the load balancer instead of working with it.

Under the new system, MCP moves to a looser, “stateless” approach to session IDs on the server side — closer to how most ordinary websites already work — which should make servers easier to maintain and, in theory, cheaper to run at scale. That gap has been a major reason companies haven’t shipped more large-scale, first-party MCP integrations despite this year’s hype around agentic AI.

Why it matters

Slower-moving infrastructure standards, not flashier model releases, may end up deciding how fast AI agents actually reach production inside real companies.