debate

Debate

Get better decisions by having agents argue different positions. Quality comes from disagreement, not consensus.

When to Use This Pattern

• A decision has multiple valid options with real tradeoffs (REST vs. GraphQL, monolith vs. microservices)
• You want to stress-test a proposal before committing
• The stakes are high enough to justify hearing both sides
• You suspect confirmation bias in the current direction

Process

Phase 1: Frame the Question

Define the debate clearly:

1. The question: What specific decision needs to be made?
2. The positions: What are the 2-3 options being considered?
3. The criteria: What matters most? (cost, speed, maintainability, user experience)
4. The context: What constraints exist? Read relevant context/ files for background.

Phase 2: Assign Positions

Spawn one agent per position. Each agent:

• Argues FOR their assigned position (not against others)
• Must provide concrete evidence: code examples, data, precedent
• Must acknowledge the strongest argument against their position
• Returns a structured brief: thesis, evidence, risks, and the one reason to pick this option

Give all agents the same context and criteria. The only difference is their assigned position.

Phase 3: Rebuttal (Optional)

For high-stakes decisions, run a second round:

1. Share each agent's brief with all other agents
2. Each agent responds to the strongest point from the opposing position
3. This round often surfaces risks that the first round missed

Phase 4: Judge

You are the judge. Do not delegate this step. Review all arguments and:

1. List the strongest points from each position
2. Identify where agents agreed (these points are likely true regardless of position)
3. Identify where they disagreed most sharply (these are the real tradeoffs)
4. Make a recommendation with clear reasoning
5. Document what you would watch for if the decision turns out wrong

Example

Task: "Should we use a monorepo or separate repos for our three services?"