Vol. IV · Multi-Agent Architectures · IV.7
Debate and Critique
Two or more models argue; a judge or a rubric decides. Use when single-pass generation is biased and you can afford the tokens.
5 min read
Doctrine
- Assign sides and a judge. Unstructured argument is entertainment.
- Debate needs a ground truth or a rubric, or it optimizes for eloquence.
- Keep the judge off the tools. Judges that can act will.
A narrow, sharp tool
Good for: safety review, architecture alternatives, translation fidelity. Bad for: anything with a side effect. Combine with evaluator-optimizer: debate proposes, rubric scores, optimizer rewrites.
Anti-patterns
- Agents debating while holding production credentials.
- No transcript review — debate that cannot be audited is theater.