Mad World

Learn

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.

Further reading

Browse example debates