Adversarial Review Improves AI Code Review with Structured Disagreement.

Early multi-agent LLM systems for coding tasks often faced diminishing returns when scaling agent numbers. While some alternatives reduced agents to passive tools, this sacrificed the benefits of agent interaction. This research explores a middle ground: minimal agentic cooperation without the overhead of large teams, introducing a protocol called Adversarial Review (AR). AR involves a main coding agent collaborating with a reviewer and a critic agent. The reviewer evaluates the code, and crucially, the critic audits this review by explicitly introducing structured disagreement before the main agent makes edits. On the LiveCodeBench, AR achieved the highest pass rate among tested methods, outperforming a five-agent baseline with only three agents. For SWE-PRBench, an initial "naive" AR exposed a false-consensus issue, but a simple prompt iteration that explicitly encouraged disagreement led to the highest F1 score. AR also showed improvements on repository-level coding tasks on SWE-bench Verified, demonstrating that effective cooperative code review can be achieved with minimal agents and structured, evidence-grounded disagreement.