How do you handle failures in complex multi-step automation workflows?

On "succeeds but produces incorrect results": the trap is treating a 200/success status as proof the output was right. Two patterns that actually catch this: (1) a structural validation step right after anything an AI produces, checking shape/required fields before it's allowed to flow further, not just "did the API call not error"; (2) for anything time-sensitive, re-checking the state you're about to act on right before you act, not just when the trigger first fired. A "wait, then act" step (approvals, delayed sends, scheduled follow-ups) can easily act on stale state if something else changed in between, e.g. the record you're about to update got deleted or already handled by a human.

For unreliable external APIs/AI steps: retry once with a short backoff, then fail loud instead of retrying silently forever, silent infinite retries are how a flaky API turns into a cost or rate-limit problem you don't notice until the bill or the ban shows up. And I'd treat "AI step returned something, but I'm not confident it's right" as its own outcome, not lumped in with "definitely wrong", route that case to a human-review queue instead of forcing a binary pass/fail.

Logging every decision (not just failures) is what actually lets you answer your own question 1 later, you need the "here's what happened and why" trail before you know there's a problem, not after.