Multi-agent

Evaluator–optimizer

One agent generates, another grades it and sends it back to improve.

2 agentsCore

How it works

1Generator produces a candidate.
2Evaluator scores it against explicit criteria and writes specific feedback.
3If it fails, the feedback goes back to the generator for another round.
4Loop until the evaluator passes it or you hit a round cap.

Use it when

There are clear evaluation criteria and iteration measurably helps — translation, writing to a spec, code against tests.

Reach for something else when

You can't articulate good criteria, or one pass is already good enough.

Where you stay in the loop

You set the bar the evaluator enforces; the two roles iterate toward it. Check that a 'pass' reflects real quality and not a criterion you specified too loosely.

In the wild

A generator drafts a translation; an evaluator checks nuance and idiom and returns notes; repeat until it passes.

Hand this to your agent

Set up a generator–evaluator loop using two roles.

GENERATOR: produce a candidate for my task.
EVALUATOR: grade it against these criteria <criteria>; give a pass/fail plus
specific fixes.

If fail, GENERATOR revises using the feedback. Alternate the two roles, show
both each round, and stop at pass or after 3 rounds.

Task: <...>

Replace the <…> placeholders, paste it into your agent, and it'll scaffold the workflow with you.