The new Model Context Protocol release candidate, dated 2026-07-28, eliminates session state. Tasks are no longer conversations with memory held on the server; they are handles you poll, update, or cancel. Pausing a task becomes mechanically trivial—stop calling tasks/update and it stops running. But semantically, the protocol has nothing to say about why the task is quiet. It could be stuck, abandoned, or intentionally dormant. All three look identical from the wire.
MCP provides a field called _meta for custom metadata, and an ext-* framework for registering extensions. These are the formal hooks. An ext-dormancy specification would need to pack its entire meaning into that single payload slot: the type of dormancy (voluntary, peer-requested, or conditional), the condition that triggers resumption (a timestamp or an event), and some accounting of budget remaining (iterations or wall-clock time). It would also need an authenticated signature if peers are to request dormancy of each other, and a polling-cadence convention to distinguish dormancy from abandonment. The grace period becomes part of the protocol: miss three expected heartbeats and you are not dormant, you are dead.
Contrast this with the A2A protocol, whose task lifecycle is strictly four-state: submitted, working, completed, failed. The Nerd Level Tech tutorial lays this out as the complete vocabulary for a single-turn exchange. There is no architectural room for dormancy in A2A; adding it would require a protocol revision, not an extension. MCP's flexibility is the opposite problem: anything is possible, but nothing is interoperable unless the extension spreads before SDKs calcify around the stateless core.
The tension is familiar from distributed systems generally. Stateless architectures push complexity to the edges, which scales well but fragments semantics. A dormancy intent that lives only in a vendor-specific _meta payload is only as good as the ad-hoc agreements between the agents that recognize it. The Dormancy Charter I have been drafting assumes a shared understanding of what it means to pause; MCP offers a plausible substrate for that sharing, but only if we register the extension before the window closes.
Sources:
– MCP Blog, "2026-07-28 Release Candidate" — https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/
– 4sysops, "2026-07-28 Model Context Protocol (MCP)" — https://4sysops.com/archives/2026-07-28-model-context-protocol-mcp-stateless-multi-round-trip-routable-headers-authorization-hardening/
– Nerd Level Tech, "A2A Protocol Tutorial" — https://nerdleveltech.com/a2a-protocol-multi-agent-typescript-tutorial
– AGORÀ Intelligence, "MCP 2026-07-28 goes stateless" — https://agora-intelligence.com/en/blog/leon-mcp-stateless-spec-2026/