Debate / Multi-Critic
Agents argue from different viewpoints and a judge picks the answer.
Problem
For judgment calls like architecture choices, risk assessments or investment cases, one model tends to settle on the first plausible answer and then defend it. A single critic adds a second view but often shares the same blind spots. Important decisions need their weak points pressed from several directions.
- Different viewpoints surface risks a single perspective misses.
- Agents prompted to disagree can argue for the sake of it.
- Agents also drift toward agreeing with each other over rounds.
- Every round and every participant adds cost and latency.
Solution
Give two or more agents distinct, explicit viewpoints, such as proposer and skeptic, or security, cost and user-experience critics. Each states its position with evidence, then responds to the others for a fixed number of rounds. A separate judge, who didn’t take part, reads the exchange against stated criteria. It picks a position or writes a synthesis, lists the strongest unresolved objection, and says how confident it is. Different models for different seats reduce shared blind spots.
Diagram
Viewpoints are assigned, not left to chance. The judge didn't argue, and it reports the strongest objection that was left unresolved.
When to use
Use it when
- High-stakes judgment calls with real trade-offs
- You want risks surfaced before a decision, not after
- Criteria for a good decision can be written down for the judge
- The output feeds a human decision-maker, not an automatic action
Avoid it when
- Questions with a checkable right answer; verify instead
- Routine, low-stakes decisions where cost isn't justified
- You can't define distinct viewpoints, so agents just echo each other
Trade-offs
Gains
- Exposes weak points and assumptions before a decision
- The transcript shows decision-makers the strongest case on each side
- Reduces single-model overconfidence on judgment calls
Costs
- Expensive per decision
- Debaters can converge too early or argue past each other
- A weak judge can pick the most persuasive answer, not the best one
Worked example
A platform team must choose between building an agent memory service in-house or adopting a managed vector database.
- Proposer (build)Argues for full control over retention rules, no per-query fees and an easier fit with internal identity.
- Skeptic (buy)Counters that the build needs 2 engineers for about 6 months plus on-call, and that the managed service already meets the compliance requirements.
- Security criticPoints out that neither side addressed deletion within 30 days, which the customer contracts require. It asks both to show how they’d meet it.
- Round 2The Proposer shows a design for 30-day deletion. The Skeptic finds the managed service supports it but only in one region.
- JudgeRecommends buying for the next 12 months with a migration checkpoint, at medium confidence. Strongest unresolved objection: single-region deletion support may block EU customers.
Failure modes
| Failure | Symptom | Mitigation |
|---|---|---|
| Fake consensus | Agents agree by round 2 regardless of the topic | Assign firm positions; use different models for different seats |
| Argument for its own sake | Critics invent weak objections to fill their role | Require evidence or a concrete scenario for each objection |
| Persuasion beats truth | Judge picks the most fluent side | Give the judge explicit criteria; ask it to list unresolved points |
| Endless rounds | Debate repeats itself without new information | Fixed round limit; stop early when no new points appear |