Agent Iterative Refinement

Agent iterative refinement

A first draft is rarely the best an agent can do, and asking it to try again produces a different draft rather than a better one. Refinement works when each round has a specific critique to act on and a standard to be measured against.

Method

1. Write the standard before the first draft. Criteria the output must meet, specific enough to check. Refining against taste produces drift rather than improvement (see agent-quality-council).
2. Critique against the criteria, one round at a time. A list of specific gaps with locations, not a general assessment.
3. Revise the draft rather than regenerating it. The instruction is to fix these things and keep the rest, or each round trades old problems for new ones.
4. Use a separate critic. Self-critique in the same context is weaker, because the model defends what it just wrote (see agent-generate-and-verify).
5. Stop on the standard, or on diminishing returns. Two rounds carry most of the gain; by the fourth you are usually paying for changes nobody wanted.
6. Watch for oscillation. A change reverted in the next round means the criteria conflict, which is a standard problem rather than a drafting one.
7. Keep every version. The third draft is sometimes worse than the second, and without history that is unrecoverable.

Boundaries

Refinement polishes; it cannot rescue a draft built on the wrong premise, where the answer is to restart with a better brief. Each round costs a full generation, so the loop needs a cap. A standard that cannot be written down cannot be refined against, and that work belongs to a human.