Persist state across multi-step workflows

Give long-running, multi-step agent tasks stable identifiers and durable state, so progress survives retries, scaling, and reconnection.

Why

Stateful sessions that collide with load balancers lose work mid-task. Anthropic's guidance on long-running agent harnesses is explicit: progress held only in an in-memory session is lost on any retry, scaling event, or reconnection. The MCP Tasks specification formalizes this with a defined lifecycle (submitted → working → input-required → completed/failed) and stable task IDs. The 12-Factor Agents framework calls for stateless reducers with durable external state — the agent's code should be able to restart from any point given the stored task state.

Do

Use durable task IDs and resumable state.

Don't

Hold critical state only in a sticky session.

Artifact

{
  "task_id": "task_88a",
  "status": "running",
  "step": 3,
  "of": 5,
  "resumable": true,
  "resume_token": "rt_...",
  "updated_at": "2026-06-04T12:00:00Z"
}