Generator-Critic / Verifier

Definition

One agent generates; another critiques, verifies, scores, or proposes revisions.

Category: Decision

When to use

Code generation + review, document drafting + editing, plan generation + verification, test repair loops.

When not to use

When the critic has no additional information or tools — it will just echo the generator's biases.

How to implement

1. The critic's prompt and tools must differ from the generator's.
2. The critic's output is structured: issue, severity, evidence, suggested fix.
3. For code tasks, the critic should run tests, linters, and diff checks where possible.
4. The generator revises against feedback for up to N iterations.

Minimal pseudocode

let draft = await generator.run(task);
for (let i = 0; i < maxIterations; i++) {
  const review = await critic.review({ task, draft });
  if (review.passed) return draft;
  draft = await generator.revise({ task, draft, review });
}
return { draft, warning: "max iterations reached" };

Recommended trace events

• generator.draft.created
• critic.review.completed
• critic.issue.found
• draft.revised

Common failure modes

• The critic only does style commentary, no fact verification.
• The loop iterates without improvement.
• Critic and generator share the same context, producing correlated errors.

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