The session ID mess that's been holding back AI agents

Next week, the Model Context Protocol gets an update that most end users will never notice, but anyone trying to run AI agents at scale should care about deeply. The change targets a fundamental headache: how MCP servers handle session IDs.

The current system is stateful. An MCP client like Claude connects to a server, says hello, and gets back a session ID it must then include in every subsequent request so the server remembers the conversation. As Arcade founder Nate Barbettini explained to us, this falls apart in real deployments where traffic is routed through a load balancer to a farm of servers that don’t share state by default. Every machine has to know about a session ID some other machine handed out. It’s a serious pain, and it fights the load balancer instead of working with it.

Arcade, a two-year-old startup that raised $60 million in June to build the infrastructure for AI agents to securely act on tools like Gmail and Slack, laid out why this has been a significant drag. Barbettini’s diagnosis matches what we’ve been hearing from engineering teams for months: most AI agents don’t fail because the models are weak. The plumbing isn’t ready. The session ID issue has meant extra work for anyone running an MCP server at scale, which is part of the reason we haven’t seen more companies ship large-scale, first-party MCP integrations despite all the noise around agentic AI this year.

The new spec, which has been public since May but rolls out formally next week, takes a stateless approach to session IDs on the server side, similar to how ordinary websites already work. The practical upshot should be easier maintenance and, in theory, cheaper operation at scale. It’s a small, technical change. But it’s exactly the kind of slow, unglamorous standards work that determines whether the agent revolution actually ships or stays stuck in demo-land.