ModelRefs / Multi-Agent Debate — Agent Pattern
Multi-Agent Debate — Agent Pattern
Two or more agents argue from different positions; a judge selects or synthesizes the verdict. A judge agent reads the transcript and produces the final answer.
Overview
Run N agents with conflicting roles (e.g., advocate, critic) over multiple rounds. A judge agent reads the transcript and produces the final answer. Improves accuracy on contested or open-ended questions.
When to use it: You need high-quality decisions on open-ended or contested questions.
Pattern details
| Pattern class | multi-agent |
|---|---|
| Difficulty | advanced |
| Autonomy | autonomous |
| Also known as | multi-agent argumentation, advocate-critic |
| Last reviewed | 2026-06-07 |
Known failure modes
- Groupthink — Sides converge on a wrong answer. Mitigation: Enforce strict role prompts; require contradictions every round.
- Judge bias — Judge favors verbosity over correctness. Mitigation: Use a rubric-driven judge prompt and rotate judges.
When not to use it
- Debating tasks with a single objective correct answer.
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Multi-Agent Debate — Agent Pattern.
Frequently asked questions
When should I use the Multi-Agent Debate agent pattern?
You need high-quality decisions on open-ended or contested questions.
What are common failure modes of Multi-Agent Debate?
Groupthink • Judge bias
Is Multi-Agent Debate production-ready?
Yes when paired with the safety controls and observability hooks documented on the pattern page.