Multi-Agent Debate
What It Does
Orchestrates structured debates between multiple LLMs to find optimal solutions to complex problems. Uses a formal Proposer → Opponent → Judge protocol with round-based argumentation. Builds consensus through structured critique, preserves minority reports when consensus isn't reached, and produces a complete debate transcript as an audit trail for human review.
Iron Laws (NEVER violate)
- Role integrity — Each model stays in its assigned role for the entire debate. Proposer advocates; Opponent critiques; Judge decides. No role-switching mid-debate.
- Evidence over rhetoric — Every argument must cite specific evidence (code, data, precedent). Rhetorical flourishes without evidence are ignored by the Judge.
- Minority report preserved — If consensus isn't reached, the minority position must be documented with equal prominence. Suppressed dissent is lost insight.
- Debate has an endpoint — Maximum rounds defined upfront. Infinite debate is analysis paralysis. After N rounds, Judge issues final ruling.
Red Flags (STOP immediately)
- Ad hominem in debate — A model critiques the other model rather than the argument → protocol violation; reset that round
- Judge bias — Judge consistently favors one model regardless of argument quality → judge rotation needed
- Circular argument — Round 3 rehashes Round 1 without new evidence → debate has converged; end it
- Proposer-opponent collusion — Models agree too quickly on a suboptimal solution → inject adversarial challenge
Common Rationalizations (self-deception)
- "One model can play all three roles" → Self-critique is weaker than adversarial critique. Separate models for honest debate.
- "Debate takes too long for simple decisions" — True. Reserve debate for high-stakes, non-obvious decisions only.
- "Consensus means the answer is correct" — Consensus can mean shared blind spots. The process improves confidence, not certainty.
When To Use
- High-stakes architectural decisions with multiple viable approaches
- Security design where adversarial thinking is critical
- Algorithm selection with tradeoffs (performance vs readability vs maintainability)
- User wants to stress-test an approach before committing
- Resolving disagreement between human team members (use AI debate as input)
Human Partner Signals (escalate to human)
- Persistent disagreement — After max rounds, models still disagree → human tiebreaker needed
- Novel domain — Debate topic is outside all models' known expertise → human domain expert needed
- Ethical dimension — Debate touches on ethics, fairness, or societal impact → human values judgment required
- Cost threshold — Debate token usage exceeds budget → human decision: continue or accept current result
Pipeline
- Frame: define the decision to be made, success criteria, and debate constraints (max rounds, time limit)
- Assign roles: select Proposer model (creative/generative strength), Opponent model (critical/analytical strength), Judge model (balanced/evaluative strength)
- Round 1 — Propose: Proposer presents solution with supporting evidence
- Round 1 — Oppose: Opponent critiques with counter-evidence and alternative risks
- Round 1 — Judge: Judge evaluates both arguments, identifies gaps, frames Round 2 question
- Round N: iterate with focused sub-questions until convergence or max rounds
- Final ruling: Judge issues decision with confidence score and key reasoning
- Archive: save complete debate transcript with minority report if applicable
Verification Checklist
Related Skills
cross-model-reviewer — Cross-model review is a simpler form of multi-perspective evaluationprovider-router — Selects which models participate in each debate rolecode-reviewer — Debate format applied to contentious code review decisionssystematic-debugging — Debate format for competing root cause hypotheses