The most important protocol in AI — the Model Context Protocol (MCP) — is quietly rolling out a critical update that could determine whether it can truly scale.

On July 21, the Arcade team published a technical deep dive explaining in detail MCP's upcoming stateless update. This update might be invisible to regular users, but for every company deploying MCP services in production, it solves a hair-pulling problem: the conflict between session IDs and load balancing. If MCP before was a toy that could only run on a single machine, after this update it can finally do real work on distributed clusters.

Why Stateful Is a Scalability Killer

To understand why this update matters, you first need to understand how MCP used to work.

The old MCP workflow went like this: a client (like Claude) connects to an MCP server and sends a "hello" — who I am, what version, what capabilities. The server replies with its own capabilities and hands back a session ID. Every subsequent request from the client carries this session ID, so the server can recognize "oh, this is the same conversation as before."

Sounds fine, right? But once you try to scale it, problems arise.

Imagine this: you have an MCP service serving millions of users, fronted by a load balancer that randomly distributes requests across dozens of servers. Here's the problem: a session ID is issued by one server, but the next request might get routed to a different server by the load balancer. That other server doesn't recognize the session ID at all. Now what?

You have a few options: one, make all servers share session state (set up a Redis cluster for sessions); two, make the load balancer do "sticky sessions" (same user always hits the same machine); three, make every server know every session ID (impossible). Every option has its own problems and costs — which is one reason many companies, despite praising MCP, have been hesitant to deploy it at scale in production.

The core of this update is: moving session state from the server to the client. The server no longer stores session state. The session ID becomes a client-managed "token" — the client brings it on every request, and the server can figure out context from the information in the token without storing anything locally. It's just like how regular websites use JWT tokens for user authentication — the server doesn't need to remember each user's session, it just needs to verify the token.

Why This Is a Milestone for AI Infrastructure

This technical detail sounds dry, but its impact might be bigger than you think.

First, MCP deployment costs will drop significantly. Previously you needed a whole session management infrastructure stack. Now servers can be completely stateless — add or remove machines at will, horizontal scaling becomes trivial. It's like going from "every waiter has to remember every customer's order" to "customers carry their own order tickets, anyone can serve them" — restaurant operating costs drop by an order of magnitude.

Second, more companies will be willing to build first-party MCP integrations. Why are most big company MCP integrations currently built by third-party developers? Because official teams think deployment and maintenance costs are too high. If going stateless makes deployment as simple as running a regular API service, then companies like Salesforce, Notion, and Figma might officially build MCP services themselves — and when that happens, both the quality and quantity of the MCP ecosystem will jump to a new level.

Third, the layered structure of agent protocols is getting clearer. MCP handles "how agents call tools," A2A handles "how agents communicate with each other" — MCP is solving the scaling problem first, while A2A is still on the way. Just like the relationship between TCP and HTTP: lower-layer protocols mature first, then upper-layer applications can explode. The MCP + A2A dual-layer protocol stack is moving from concept to usability.

Arcade itself is worth watching. Founded just two years ago, it raised $60 million in June, and its entire business is built on the proposition of "making AI agents actually work inside enterprises." Its founder coming out to explain the MCP update also represents, in a way, the real demand for enterprise-grade agent infrastructure.

Many AI breakthroughs are loud — new model launches, new SOTA records, new funding announcements. But what really determines whether an industry can land is often these quiet infrastructure improvements. Just like in the history of the internet, the WWW didn't spread because of any one hot website — it spread because underlying protocols like TCP/IP, HTTP, and HTML matured one by one.

MCP going stateless is exactly that kind of unassuming but essential brick in the wall.

See you tomorrow.