Motivation
Many agent tasks require quality refinement, not just validation. Contract drafting, report writing, policy review, and code generation benefit from structured critique cycles rather than single-pass generation.
The Critic-Actor Pattern separates generation from evaluation, creating a feedback loop that converges on acceptable quality. Unlike the Validation Loop (which tests correctness), the Critic-Actor pattern improves quality through structured critique.
Structure
- generate() — Actor produces the first draft
- critique() — Critic evaluates the draft and returns structured feedback
- refine() — Actor revises the draft based on the critique
- should_accept() — Critic decides whether the refined output is acceptable
- Repeat or finalize / escalate