When to reach for it
- The problem allows for controversial assessments.
- Counterarguments help expose logical flaws.
- Decisions require rigorous verification.
Multi-Agent Debate.
Argue the answer. An arbiter decides.
Agents represent different positions and critically negotiate outcomes before a final decision is synthesized.
When it backfires
- Facts are easily verifiable.
- Debate would create artificial conflicts.
- Latency budgets are tight.
The tradeoff
Rigorous verification of difficult decisions is achieved against increased effort and the risk of over-arguing.
What it actually does.
Agents argue, then converge on a consensus answer.
Two ways this pattern will hurt you.
Debaters confuse verbosity with correctness
The arbiter picks the longest or most confident argument, not the most accurate one. The debate becomes a rhetoric contest.
Fix · Score arguments against verifiable criteria (citations, test results) rather than persuasive style.
Artificial conflict on settled facts
The debate format forces agents to argue against an established truth. Time is wasted manufacturing false controversies.
Fix · Run a fact-check pass before debate. Skip debate when all agents agree on the factual premise.
Threat exposure
What this pattern exposes
Adopting this pattern opens these attack surfaces. Each links to its entry in the threat model.
Full threat model →Where Multi-Agent Debate is native.
Microsoft Agent Frameworkmulti-agent debateNative