E5 · Group chat and debate

Principle

The agents share one discussion thread. A moderator (chat manager) hands out turns; each one proposes, critiques and revises until they converge. In its academic form, the debate has several instances of one model revise their answers over several rounds.

The problem it solves

A single model can be wrong with nothing to warn it; the hope is that crossed viewpoints produce a better answer or a consensus.

When to use it

• Brainstorming, consensus building, quality or compliance review by several “experts”, with an objective stopping criterion.
• The models are heterogeneous: that is the only case where debate adds value of its own.

When to avoid it

• At equal cost, prefer sampling and voting (B4): voting explains most of the gain attributed to debate.
• Real-time use, no objective stopping criterion, or more than three agents.

Common pitfalls

The anti-patterns that misuse this pattern, best established first. Each sheet describes the symptoms, the public incidents and the remedies.

Loop, stopping and orchestration

• AP35 · Unbounded agent loop — Nothing in the code stops the agent: no turn cap, no budget, no deadline.
• AP41 · Conformist debate — Agents debate to correct one another and end up agreeing on the most confident answer, even a wrong one.

All anti-patterns

Synonyms by vendor

• Group chat, roundtable, collaborative, multiagent debate, council (Microsoft)
• Multiagent debate (Du et al. (ICML 2024))
• Debate-based and role-based cooperation (CSIRO Data61)
• Conversation patterns (CAMEL, ChatDev, AutoGen)

Neighbouring patterns

• E2 · Task-ledger orchestrator (Magentic)
• E6 · Mixture-of-agents (layered)
• E7 · Swarm and peer network
• B4 · Sampling and voting

Sources

• Improving Factuality and Reasoning in Language Models through Multiagent Debate — Du et al., ICML 2024, arXiv, 2023-05
• Should we be going MAD? — Smit et al., ICML 2024, arXiv, 2023-11
• Stop Overvaluing Multi-Agent Debate — Zhang et al., arXiv, 2025-02
• Debate or Vote — Choi, Zhu, Li, NeurIPS 2025, arXiv, 2025-08
• AI agent orchestration patterns — Microsoft, Azure Architecture Center, 2026-02-12