The Debate Pattern

The debate pattern is the most compelling reason to use agent teams over sub-agents. Spawn two or more agents with deliberately different perspectives and let them argue through a shared message board until they reach consensus. The pattern emerged from security auditing: instead of having one agent scan a codebase and produce a report, spawn multiple scanner agents that then hand off to 'debate agents' who argue about which findings are real vulnerabilities versus false positives. The debate surface tensions and nuances that a single-pass review consistently misses.

1. 1. Define Opposing Positions Spawn Agent A as the advocate (e.g., 'this code is secure') and Agent B as the challenger (e.g., 'find every possible vulnerability'). Give each agent a clear perspective to defend.
2. 2. Provide a Shared Artifact Both agents need access to the same codebase, document, or data set. Point them to the same files or provide the same context in their spawn prompts.
3. 3. Open the Debate Channel Agent teams communicate via a shared scratch pad. Agent A posts its analysis, Agent B responds with counterpoints, and the cycle continues. The team lead monitors but doesn't intervene unless the debate stalls.
4. 4. Define Convergence Criteria The debate should have explicit exit conditions: 'Debate until both agents agree on the top 5 findings ranked by severity' or 'Continue for a maximum of 3 rounds, then the lead synthesizes.' Without this, agents will debate forever.
5. 5. Synthesize and Act The team lead collects the debate transcript, extracts consensus points and unresolved disagreements, and presents a final recommendation. Unresolved items get flagged for human review.