Creativity · Agent Protocol

Multi-Agent Debate Pattern

Multi-agent debate is a reasoning pattern in which two or more LLM agents are assigned opposing or complementary positions and argue the question out in rounds. A judge agent — or the original user — reads the transcript and decides. Research from DeepMind, MIT, and others has shown that debate often improves accuracy on hard reasoning tasks by surfacing errors that single-chain reasoning misses.

Protocol facts

Sponsor
open research pattern
Status
stable
Spec
https://arxiv.org/abs/2305.14325
Interop with
LangGraph, AutoGen, CrewAI, A2A

Frequently asked questions

Why does debate improve accuracy?

An agent arguing the opposing side has an incentive to find flaws in the first agent's reasoning. Errors that a single chain-of-thought would miss get flagged, and the judge sees both sides before deciding.

How many agents do you need?

Two debaters plus a judge is the canonical minimum. Some systems use more debaters with different roles — e.g., optimist, skeptic, domain expert — but returns diminish after three or four voices.

Is debate worth the extra cost?

Often yes for high-stakes reasoning — math, legal, medical — where a wrong answer is expensive. For most chat or retrieval tasks, simpler patterns win on cost and latency.

Sources

  1. Du et al. — Improving Factuality via Multiagent Debate — accessed 2026-04-20