Three releases landed inside seven days — Vercel's AI SDK 7, VS Code 1.129, and the release-candidate MCP 2026-07-28 stateless core. They come from three companies solving three different problems, and if you read the changelogs side by side they are making the same move: taking the state that used to live inside a single request, process, or connection, and pushing it somewhere that outlives the request ending.
The short version: the agent stack has stopped pretending an agent run is one HTTP call. AI SDK 7 persists run state to durable storage between steps; VS Code moves the agent into its own crash-isolated process; MCP deletes the session so state has to travel as explicit handles. The assumption underneath all three is that your agent runs for minutes, gets interrupted, and has to survive it. Design for that, or keep losing runs to a deploy.
What each one actually shipped#
AI SDK 7 — WorkflowAgent. The headline primitive persists an agent's execution state to durable storage between steps, so a long-running agent survives deploys, process restarts, interruptions, and — the one that bites every founder — a delayed human approval. That's Vercel folding durable execution into the SDK itself, instead of making you reach for Temporal, Inngest, or a hand-rolled checkpointer. AI SDK 7 also ships first-class timeout budgets (total, per-step, per-chunk, per-tool), which is the same instinct from the other direction: a long run needs bounds you can set explicitly.
VS Code 1.129 — the agent host. Shipped July 15 (with a 1.129.1 fix on July 17), it runs agents in a dedicated host process built on an Agent Host Protocol, running harnesses like Copilot, Claude, and Codex outside the editor window. Two consequences matter: a crashed agent no longer takes your editor down with it, and a single agent session can attach to multiple windows at once. The agent stopped being a thing that lives and dies inside one editor window. (We covered the full 1.129 agent update here.)
MCP 2026-07-28 — the stateless core. The spec locks July 28, seven days out. It removes the session handshake so any server node can serve any request — which only works if the session state stops living in the connection and starts traveling as explicit handles. It is the exact same refactor as the other two, expressed as a protocol instead of an SDK or an editor.
Why three teams made the same move at once#
Because the thing they're all building for changed shape. An agent in 2024 was a request: prompt in, tool calls, answer out, done in seconds. An agent in 2026 waits on a human, runs a background job, pauses for a build, resumes tomorrow. The moment a run can outlive the process that started it, keeping state inside that process becomes a bug. So the state has to move: to durable storage (AI SDK), to a separate process (VS Code), to an explicit handle (MCP).
Put plainly: request-lifetime and state-lifetime used to be the same thing, and this week three tools finally pulled them apart. That's the non-obvious part. It's not three features. It's one architectural admission, shipped in triplicate.
What a founder does Monday#
You don't need all three. You need to stop modeling an agent run as a single request in your own code, because the tools underneath you already have:
- Persist run state outside the process. Whether it's a WorkflowAgent, a durable-execution engine, or your own checkpoint table, the run has to be reconstructable after a restart. Treat resume-after-interruption as the normal path, not the exception you'll add later.
- Give the agent an identity that outlives a connection. If you talk to MCP servers, the stateless migration is due in a week — and its whole point is that the identity and state can't ride the socket anymore.
- Isolate the process. VS Code's lesson generalizes: don't run the agent in the same process as the thing you can't afford to crash. A separate worker is cheap insurance now that a run lasts minutes.
The stack spent this week telling you agents are long-running. The founders who ship reliable ones are the ones who believe it before their first deploy eats a half-finished run.
