Coding agents
Write code, run tests, read failures, fix, repeat. A verifiable stop condition (tests pass) makes long-running "while not done" loops safe. The most mature use case by far.
The pattern that fits depends on the shape of the work, the cost of a bad output, and whether "done" is checkable.
Write code, run tests, read failures, fix, repeat. A verifiable stop condition (tests pass) makes long-running "while not done" loops safe. The most mature use case by far.
Explore broadly, narrow to sources, cross-check, cite. A supervisor splits the search space; an evaluator loop grades coverage before the report ships.
Long-form writing where a second pass reliably beats the first. Generate → critique → revise, capped at two or three cycles before diminishing returns set in.
Scheduled loops that check a signal, decide whether action is needed, and page a human or fire a webhook. Idempotency is the whole ballgame.
Pull structured data from messy sources with schema-conforming outputs. Retry on schema failures; sample-verify against a held-out set.
Image, audio, or video where quality is judged, not measured. Loop with a taste-critic model or human sign-off; cap iterations by budget. Taste stays in the gold ring.