Multi-agent debate explained
Multi-agent debate is a setup where two or more language models take turns arguing a question so disagreement and critique are visible — not a single model answering alone.
Multi-agent debate is a setup where two or more language models take turns arguing a question so disagreement and critique are visible — not a single model answering alone.
The basic idea
Instead of one assistant producing a polished answer, several models (often from different labs) respond to the same prompt and to each other. The hope is that peer critique catches errors, surfaces tradeoffs, and makes uncertainty honest.
Mad World turns that idea into a product room: Discussion for open multi-turn debate, Council for independent answers plus ranking, and a judge that tracks agreement versus contested points.
When it helps
- Questions with genuine tradeoffs (investing, product, policy).
- Claims that need sources, not vibes.
- Cases where one model’s confidence is the risk.
When it doesn’t
Homogeneous clones revising under peer pressure can add noise or corrupt a correct first answer. See when multi-agent debate fails and the research roundup.
Try it
Browse example debates or start a room on Mad World.