Refinement Loop / Evaluator-Optimizer

Definition

Iterate generation → evaluation → revision until an exit condition is met or the budget is exhausted.

Category: Decision

When to use

Test-driven code repair, document polishing, prompt optimization, plan iteration, quality gates.

When not to use

When there's no clear evaluation criterion, no defined exit condition, or budget is tight.

How to implement

1. Define exit conditions up front: tests passing, score above threshold, human approval, no new issues.
2. Each round only modifies explicit failure points — avoid oscillation.
3. Log per-round diff, score, feedback, and cost.
4. On max-rounds, return current state and unresolved issues.

Minimal pseudocode

for (let round = 1; round <= maxRounds; round++) {
  const output = await worker.run(state);
  const evalResult = await evaluator.run(output);
  logRound(round, output, evalResult);
  if (evalResult.pass) return output;
  state = applyFeedback(state, evalResult.feedback);
}
return { status: "incomplete", state };

Recommended trace events

• loop.round.started
• loop.evaluation.completed
• loop.exit.pass
• loop.exit.budget_exceeded

Common failure modes

• Vague exit conditions create infinite loops.
• Each round introduces new bugs.
• The system optimizes the evaluator's score, not the real goal.

Implementation checklist

• Input/output schemas defined.
• Each agent's permission boundary defined.
• Every agent call carries a run id / trace id.
• Failure, timeout, cancel, and retry strategies defined.
• Context passed is the minimum required, not the full history.
• High-risk actions are gated by approval or a verifier.

References

• Google ADK patterns
• Google architecture patterns