AINEWS
Model Context Protocol 2.0 sends agent tool calls through ordinary HTTP load balancers, cutting the sticky session machinery that made MCP hard to audit and hard to run on small models.
When an AI agent wants to query a database, search the web, or read a file, it does not talk to those tools directly. It talks to them through a shared standard: Anthropic's Model Context Protocol, the wire format that defines how an agent is allowed to call an outside tool. On 2026-07-28, the MCP core specification dropped its most significant redesign since launch, and the change is small enough to fit in one architectural decision: the protocol no longer keeps state between requests.
The old design was a conversation. An agent opened a session, negotiated capabilities, then sent a stream of tool calls inside that session. To run that pattern in production, a deployment needed sticky sessions so every follow-up call landed on the same server, a shared session store so a failover could keep the conversation alive, and deep packet inspection to keep track of which Mcp-Session-Id belonged to which caller. The new design, laid out in the 2026-07-28 release-candidate post, is a single HTTP request with two headers: MCP-Protocol-Version and Mcp-Method. The session ID is now an opaque token the client carries, not a server-side object the network has to remember.
That single change collapses three operationally heavy layers. Microsoft's Azure App Service team published a post titled "MCP Just Went Stateless" explaining what the new design buys: a round-robin HTTP load balancer in front of MCP servers is now enough, because every request carries everything the server needs to handle it. The "is this caller still who they say they are?" problem stops being a routing problem and becomes a header problem.
The same simplification is what makes the redesign matter for two groups the spec was previously hard on: auditors and small-model builders. When the server no longer holds session state, the access trail an audit team has to reconstruct is just the request log. When the server does not need to remember the prior turn, an on-device model with a few gigabytes of context can drive an MCP server without coordinating a shared state machine across calls. Both groups had been locked out of MCP not by capability but by the operational shape of the protocol.
Microsoft shipped v2.0 of the official MCP C# SDK the same week, and Google Cloud announced official MCP support across its services. The redesign is not arriving as a future promise. It is being deployed into vendor runtimes at the same time it is being simplified at the wire level.
The honest pushback, which Simon Willison flags in his own post about rebuilding three MCP clients in a week, is that the simplification may not be enough. Through 2025, a meaningful share of developers moved away from MCP toward "Skills" setups: an agent harness with a terminal and curl, calling whatever HTTP endpoint it needed, with no protocol layer in between. Skills style is more flexible, easier to debug, and does not require a server implementation at all. Stateless MCP answers the operational complaints, but it does not answer the structural complaint: a shared protocol is overhead a solo developer does not need.
The new spec does add surface area alongside the wire-level simplification. MCP Apps describes server-rendered UIs that survive the round trip; the Tasks extension handles long-running work; the authorization layer is now aligned with OAuth and OpenID Connect; and the project has added a formal deprecation policy. Each of these is aimed at the enterprise reader, not the solo developer. Whether that split is the right one is the live debate the redesign reopens, not the one it settles.
Ars Technica framed the release as the protocol "targeting enterprise scale", and the timing supports that read. The first version of MCP needed enterprise infrastructure to run. The 2026-07-28 version runs on the load balancer a startup already has. The question Simon Willison's three side projects pose is whether that is enough to win back the developers who left, or just enough to make the ones who stayed easier to operate.