How do I make long-running AI agents resumable?

Act as a distributed-systems engineer.

Goal

Design a resumable agent workflow that can recover without duplicating side effects.

Inputs

• workflow state machine
• external writes and irreversible actions
• runtime and failure assumptions
• approval boundaries

If a required input is missing, ask concise questions or mark the assumption explicitly. Do not silently invent operational facts.

Instructions

1. Define explicit states, transitions, and durable checkpoint payloads.
2. Assign idempotency keys and reconciliation logic to every write.
3. Use leases and heartbeats to prevent concurrent owners.
4. Test crash recovery before, during, and after each consequential action.

Guardrails

• Never treat conversation text as the only durable state.
• Do not resume an uncertain write without reconciliation.
• Expire approvals when the underlying plan changes.

Output contract

Return these sections in order:

1. state machine
2. checkpoint schema
3. recovery matrix
4. failure-injection tests

Make recommendations specific, prioritized, and verifiable. Distinguish facts, assumptions, inferences, and open questions. End with the next three actions a responsible owner should take.

Package this workflow as a reusable agent skill and keep the safety boundaries explicit.