Quality Through Review

The single biggest quality improvement in agent teams comes from adding a reviewer. A dedicated critic agent that evaluates output against specific criteria catches errors, weak arguments, missing information, and formatting problems that the original author agent will not see. This is not because the author agent is bad — it is because reviewing your own work is structurally harder than reviewing someone else's.

The review loop works like this. The worker agent produces a first draft. The critic agent evaluates it on specific dimensions — accuracy, completeness, clarity, formatting — and either approves it or sends it back with specific feedback. The worker revises based on the feedback and resubmits. This cycle repeats until the critic approves or the maximum number of rounds is reached.

1. 1. Define Review Criteria Tell the critic exactly what to evaluate. "Is this good?" is useless. "Score accuracy 1-10, completeness 1-10, clarity 1-10. Approve if all scores are 7 or higher. Otherwise, return specific improvement instructions" is actionable.
2. 2. Set a Round Limit Without a maximum number of review rounds, the loop can continue indefinitely. Three rounds is the recommendation. If the output is not good enough after three rounds of revision, the problem is in the prompt or the role design, not in the number of iterations.
3. 3. Define "Good Enough" The critic needs a clear approval threshold. A numeric score with a minimum bar works well. This prevents both premature approval (critic is too lenient) and endless loops (critic is too strict).