Adversarial Review Protocol Outperforms Large Multi-Agent Teams in Code Review

The Adversarial Review (AR) protocol uses a three-agent setup—coder, reviewer, and critic—to implement minimal cooperative disagreement. AR achieved the highest pass rate on LiveCodeBench among tested methods, outperforming a five-agent baseline.