adversarial-agent-review

One AI agent systematically attacks and challenges the work of another AI agent to find errors, gaps, and false assumptions before they reach production. Based on DeepMind research showing adversarial review improves output quality by 17x compared to single-agent review. Use this skill when an AI agent produces any output that will be acted upon: code, architecture decisions, business analysis, security assessments, legal summaries, financial models. Triggers on: "review this output", "check this analysis", "verify this plan", "second opinion on", "challenge this", "find problems with", "red team this", "stress test this decision", "what could go wrong", "agent wrote this".