The Model Context Protocol (MCP) is basically the plumbing of the agent era. It’s the standard that lets an AI model securely reach external data sources and services — your calendar, your database, your internal tools. Without MCP, every connection would have to be wired up by hand. Anthropic invented the protocol, and by now OpenAI, Google, and dozens of platforms rely on it.
Next week MCP gets an update end users will never see — but one that could shape how the whole ecosystem develops.
The session ID problem
Here’s how MCP works today. The first time a client like Claude connects to a server, it sends a «hello» — here I am, here’s my version, here’s what I can do. The server replies and hands back a session ID. From then on, the client sends that ID with every request so the server knows: this is the same conversation as five seconds ago.
Sounds harmless. It becomes a nightmare the moment you go big. 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 machine happens to be free — sometimes in a different region. Suddenly every one of those machines has to know about a session ID that some completely different machine handed out. Not impossible, but a serious pain. It fights the load balancer instead of working with it.
According to Russell Brandom (TechCrunch), that’s one of the reasons we’ve seen so few large-scale, first-party MCP integrations this year despite all the agent hype. Running things at scale was simply too much of a chore.
The fix: just forget
The new version flips it around. Instead of the server remembering every session, MCP moves to a looser, stateless approach — the way ordinary websites have worked for ages. The server no longer needs to remember who you are. That makes it much easier to maintain and, in theory, cheaper to scale. The official spec has been public since May, and the release is planned for next week. The clearest explanation of the change, by the way, came from the folks at Arcade.
My take
I find this news refreshing for an unusual reason: it’s a reminder that not every part of AI development is racing at breakneck speed. While model training charges ahead, a lot of the infrastructure those models depend on moves at the slow pace of standards-body consensus. It’s happening — just a little slower.
And yet that’s exactly the more important story. Models keep getting smarter, but the bottleneck for real adoption rarely sits with the model. It sits with unglamorous things like session IDs behind a load balancer. If you want agents to genuinely land in the enterprise, these are the updates worth watching — even when nobody’s throwing a party for them.
Sources: