The Model Context Protocol, the standard that lets AI models plug into your calendar, your database, and your internal tools, is getting a meaningful upgrade next week. According to TechCrunch AI, the change lands quietly at the infrastructure layer, so end users won’t see it. But it targets one of the real reasons big companies haven’t shipped large-scale AI agent integrations this year, even with all the agentic hype.
The clearest breakdown came Monday from Arcade, a two-year-old startup that TechCrunch AI reports has built its whole business around getting AI agents to actually work inside real companies, connecting securely to tools like Gmail, Slack, and Salesforce. Arcade raised $60 million in June on a simple thesis: most agents don’t fail because the models are weak. They fail because the plumbing around them isn’t ready.
What actually changed
The fix is about session IDs. Those are the small tokens a server uses to remember that this request is part of the same conversation as the one five seconds ago.
Here’s how it works today, as Arcade founder Nate Barbettini describes it. When a client like Claude first connects to an MCP server, it sends a hello with its version and capabilities. The server replies and hands back a session ID. From then on, the client attaches that ID to every request so the server knows who it’s talking to. If the ID expires, the client has to notice, ask for a new one, and continue.
That’s fine on one machine. It breaks down at scale.
Picture a server running for millions of users behind a load balancer. The load balancer’s whole job is to send each request to whatever machine is free, sometimes in another region. Suddenly every machine needs to know about a session ID that some other machine handed out. As Barbettini puts it, “It’s not impossible, but it’s a serious pain, and it fights the load balancer instead of working with it.”
Why it matters
The old design assumes one server remembers you. Real companies spread traffic across dozens of servers that don’t talk to each other by default. So today’s MCP servers do extra work just to keep track of who’s who, and that friction has been a genuine headache for anyone running MCP at scale.
The new version takes a looser, stateless approach to session IDs on the server side, closer to how ordinary websites already operate. The payoff, in theory:
- Easier to maintain, since machines don’t have to share session state
- Cheaper to run at scale, because it works with the load balancer instead of against it
- Fewer barriers to shipping first-party, large-scale MCP integrations
The official spec has been public since May. Next week’s rollout is when it starts mattering in practice.
What stands out here is the reminder that not every part of AI moves at breakneck speed. Model training races ahead. The infrastructure those models depend on still moves at the pace of standards-body consensus, which is slower and less glamorous. It’s happening, just on a different clock.
What to expect
If you’re building or running agents, this is the kind of change that quietly widens what’s possible. The models were rarely the bottleneck. The connective tissue was. A stateless session model removes one of the bigger reasons teams hesitated to put agents in front of millions of users.
Watch for more first-party MCP integrations from larger companies once this ships and tooling catches up. The hype around agentic AI has run well ahead of the plumbing all year. This is the plumbing catching up. Full details are in the original TechCrunch AI report.