Many minds on one problem
Multi agent debate has several agent instances independently answer a question, then read each other answers and argue. Over a few rounds they critique, defend, and revise, often converging on a more accurate result than any single pass.
The debate protocol
- Propose: each agent gives an answer with reasoning.
- Exchange: every agent reads the others responses.
- Critique and revise: each points out flaws in the others and updates its own view.
- Converge: after a fixed number of rounds, take the majority or a synthesized answer.
Why it can help
- Independent errors often differ, so seeing peers exposes mistakes.
- Defending a claim forces stronger justification.
- A diverse set of starting answers covers more of the solution space.
The failure modes
- Herding: agents may converge on a confident but wrong answer, amplifying a shared bias.
- Cost: many agents over many rounds is expensive.
- Sycophancy: an agent may cave to others rather than hold a correct minority view.
Debate works best when the agents are genuinely diverse and the question has a checkable answer, so wrong claims get exposed rather than reinforced.
Key idea
Multi agent debate has several agents propose critique and revise across rounds before converging, exposing independent errors for higher accuracy, but it risks herding and cost and needs genuine diversity to avoid amplifying a shared bias.