Deleting a session ID does not delete state. It changes who must own it.
That is the important shift in the Model Context Protocol specification scheduled for July 28. MCP's new core removes the initialization handshake and protocol-level sessions that tied clients to server instances. Any request can instead land on any compatible server behind an ordinary load balancer.
The thesis: stateless MCP should simplify transport, but production teams will still have to make three kinds of state explicit—business state, compatibility state, and control state.
What Is Actually Changing
The official release candidate calls this MCP's largest revision since launch. Six specification proposals combine to remove the `initialize` handshake and `Mcp-Session-Id`. New method headers make traffic easier to route, cache metadata can reduce repeated discovery calls, and W3C Trace Context can follow a tool call across services.
This does not make agent workflows stateless. A shopping tool can still create a basket ID; the model then passes that handle into later calls. The difference is architectural: application state becomes a visible argument instead of hidden transport metadata.
The change is also breaking. A client built for the 2025-11-25 protocol cannot automatically communicate with a server expecting 2026-07-28. Version negotiation helps, but somebody must carry the transition.
The Three-State Model
Operators should separate three responsibilities that session machinery once blurred together.
1. Business state
Long-running work, browser sessions, carts, and approvals still need durable state. Give each object an explicit handle, define its owner and lifetime, and make replay behavior predictable. If a retry reaches a different server, the result should not depend on which process remembers the conversation.
2. Compatibility state
For a transition period, clients and servers will speak different protocol versions. Track supported versions as inventory, not tribal knowledge. Test every important client-server pair, publish migration status, and decide where translation belongs.
3. Control state
Authorization, policy, consent, rate limits, and audit evidence cannot disappear with the handshake. Bind controls to identities and actions, then enforce them on every call. Transport version should not determine whether a dangerous tool action is reviewed.
Why Arcade Is Worth Watching
Arcade.dev, a private agent-infrastructure company, says its gateway already supports both the previous stateful MCP version and the new stateless one. Its pitch is that version translation belongs in a controlled runtime while authorization, security policy, and governance remain stable above the protocol.
That positioning has financial backing. Arcade announced a $60 million Series A in June, bringing stated total funding to $72 million. TechCrunch connects the funding to a broader enterprise problem: models may be capable, but the infrastructure around agent actions is still difficult to operate.
The funding is a market signal, not product validation. The reviewed sources do not establish Arcade's customer count, gateway throughput, migration success rate, or cost savings. Those are the metrics buyers should request.
A Six-Step Migration Playbook
1. Inventory every MCP client, server, SDK, gateway, and protocol version.
2. Add conformance tests before changing production traffic.
3. Support both versions at one controlled boundary where necessary.
4. Replace implicit session assumptions with explicit, scoped handles.
5. Trace identity, policy decisions, tool calls, retries, and outcomes end to end.
6. Retire the old path only after error rates and unsupported-version traffic reach an agreed threshold.
Measure compatibility failures, retry duplication, stale-handle errors, policy mismatches, and trace completeness. A migration is not complete because requests return 200. It is complete when mixed-version traffic is predictable, governed, and observable.
The Takeaway
MCP's stateless core is a useful simplification because it forces hidden infrastructure decisions into the open. It removes one source of operational memory while making business state, compatibility, and control design impossible to ignore.
The protocol will absorb less complexity. Production teams—and companies like Arcade—are competing to absorb the rest.
Sources
- Model Context Protocol Blog, “The 2026-07-28 MCP Specification Release Candidate” (2026-05-21): https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/
- Arcade.dev, “MCP '26: MCP Is Going Stateless. Here's What That Means” (2026-07-20): https://www.arcade.dev/blog/mcp-going-stateless/
- TechCrunch, “AI's most important protocol is getting a little bit easier to use” (2026-07-20): https://techcrunch.com/2026/07/20/ais-most-important-protocol-is-getting-a-little-bit-easier-to-use/
- Arcade.dev, “We saw the action layer coming. Now we're going to own it.” (2026-06-12): https://www.arcade.dev/blog/arcade-series-a/