← Back to Docs
Recipe

Critic Agent Design

A self-correcting agent loop that generates output, critiques it against a rubric, and iterates until quality thresholds are met.

Core Loop

1. Generate → initial draft

2. Critique → score against rubric

3. Revise → apply feedback

4. Repeat → until score ≥ threshold

Rubric Dimensions

  • Accuracy — factual correctness, no hallucinations
  • Completeness — all requirements addressed
  • Clarity — unambiguous, well-structured
  • Conciseness — no fluff, minimal token waste

Guardrails

Max iterations

5

Score threshold

0.85

Token budget

8K / turn

Fallback

best-so-far