MCP stateless GA is 18 days out: the week-three handle migration sprint I run before July 28

Introduction

I have watched three teams hit week 3 of the MCP stateless migration and discover their checkout flow still assumed Mcp-Session-Id would route the second tools/call to the same pod. Stateless MCP is not "delete Redis and hope." It is "move state into handles your server owns and test round-robin before GA."

July 28 is 18 days out. If you finished week 1 (grep for session assumptions) and week 2 (SDK RC bumps), week 3 is the handle sprint. This post is the checklist I run on engagements when GA is close enough that procrastination becomes an outage.

What week 3 actually means

• Every multi-step workflow gets an explicit handle schema documented in the tool definition.
• Server mints the handle on step 1; client passes it on steps 2–N as a required argument.
• Handles are bound to oauth_subject or tenant_id; cross-tenant replay must fail closed.
• Round-robin load tests run without sticky sessions on the gateway.
• List/read endpoints emit ttlMs and cacheScope before you rely on client-side caching.

The week-three sprint checklist

1. 01
   Document handle schemas in tool definitions

   Add handle fields to inputSchema with format and description. Clients should not guess handle shape from error messages. If your MCP server uses OpenAPI-style tools, publish handle lifecycle in the tool description "when to use" block.
2. 02
   Mint handles server-side on step 1 only

   Never trust client-generated UUIDs for stateful workflows unless you also verify ownership on every subsequent call. Bind handle creation to the authenticated principal from OAuth or EMA.
3. 03
   Run round-robin without sticky sessions

   Point two or more server instances behind a plain load balancer. Execute a three-step tool flow ten times. Any 502 or empty response on step 2 means state still lives in protocol memory or pod-local RAM.
4. 04
   Verify cacheScope before enabling client caches

   Hit tools/list as two different OAuth identities. cacheScope must prevent cross-user bleed. Wrong scope is a confidentiality bug, not a performance tweak.
5. 05
   Pair with header security audit

   Week 3 handle work often touches gateway routing. Re-run the MCP header security checklist so Mcp-Method paths do not accidentally carry secrets into proxy logs.

What not to do in week 3

• Deleting Redis before handles work in staging.
• Treating handles as optional "nice to have" on checkout flows.
• Running EMA rollout and stateless cutover in the same maintenance window.
• Skipping load tests because "we only have one instance in prod."
• Assuming deprecated Roots/Sampling disappear on GA day (12-month window).

How this connects to observability

Add handle_id, Mcp-Method, cache hit/miss, and oauth_subject as required dimensions in the agent observability stack. Week 3 failures show up as mysterious step-2 errors until you can trace which instance served which call.

If you run MCP through a governance gateway, confirm handle minting events appear in audit exports. Stateless routing makes session IDs disappear; handles become your correlation key.

Conclusion

Week 3 is the difference between stateless MCP on paper and stateless MCP in production. Mint handles, test round-robin, verify cache scope, audit headers. July 28 should be boring. Boring is the goal.

Sources: MCP 2026-07-28 release candidate at https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/; SEP-2567 explicit state handles; SEP-2575 stateless transport.