adversarial-team-pattern

# Adversarial Team Pattern Multi-agent analysis using opposing mandates — not consensus. Agents challenge each other rather than reinforce each other. **Core principle:** consensus from multiple agents = one agent with extra compute. Value comes from *productive disagreement*. --- ## When to use - High-stakes decisions (architecture, security, major refactor) - Analysis where confirmation bias is a real risk - Any time you'd normally just "get a second opinion" (make it an adversarial second opinion) - Code review of critical paths - Research where you want to surface what could go wrong, not just what looks right Do NOT use for: simple lookups, single-file changes, mechanical tasks with a clear right answer. --- ## Team structure Spawn these roles in a single message (parallel, `run_in_background: true`): ``` ROLE A — Advocate Mandate: Build the strongest case FOR the proposal. Find evidence that supports it. Assume it will work. Deliver: 3–5 concrete supporting arguments with evidence. ROLE B — Challenger Mandate: Find everything wrong with the proposal. Assume it will fail. Find the weakest points. Deliver: 3–5 specific failure modes or risks with evidence. ROLE C — Constraint Finder Mandate: Find what's missing, what's not accounted for, what the other two missed. Be the skeptic of both sides. Deliver: unknowns, edge cases, unstated assumptions. ROLE D — Synthesizer (run AFTER A, B, C complete) Mandate: Read all three reports. Produce a...