When multi-agent debate makes things worse
Multi-agent debate fails when identical models revise under peer pressure, when sycophancy overrides a correct answer, or when conversation replaces selection and ranking.
Multi-agent debate fails when identical models revise under peer pressure, when sycophancy overrides a correct answer, or when conversation replaces selection and ranking.
Failure modes
- Clones of one model “debating” each other.
- Forced agreement or revision toward the loudest peer.
- Weak baselines that make debate look better than it is.
- No tools — eloquence without evidence.
How to avoid them on Mad World
- Pick models from different labs.
- Use Council when you need independent answers first.
- Turn search on for factual claims.
- Read the judge’s contested list; don’t stop at a confident closer.