How do you resume an agent's work after a crash?
Answer out loud or in your head first, then compare.
Answer
I read the saved state and establish the last confirmed step. I don't rely solely on the model's message. If it's unclear whether a remote write succeeded, I first check the target system by the operation's identifier. Only then do I decide whether to retry.
With many steps, saved results and intermediate states are useful. Some operations can be reversed, but not every one: a sent email or a disclosed document can't be undone with an ordinary rollback. So recovery after failure has to be planned in advance. A good test deliberately stops the program after a write but before confirmation, and checks that resuming doesn't duplicate the effect.
Why it matters: A system's autonomy requires control over what happens after an error.
Possible follow-ups: When do you need a manual decision instead of an automatic retry?
How did it go? Be honest — “Not yet” puts the question on your review list.
Start with
- H-P01 · How does an AI agent differ from a plain model call? — Agents and workflows
- H-P02 · What is a workflow, i.e. a fixed sequence of work? — Agents and workflows
Further reading
- AWS Prescriptive Guidance: Transactional outbox(opens in a new tab)Docs; accessed 22.10.2026. The dual-write problem and writing an event in the same transaction.