Adversarial Review: Cooperative Code Review through Structured Disagreement

Early multi-agent LLM systems often used role-separated teams that debate or aggregate independent outputs, yet scaling agent count yields diminishing returns on repository-level coding tasks. In comparison, recent work has shifted towards a main-agent-plus-subagents paradigm, where subagents act like tools rather than independent collaborators. We study whether this paradigm can support a middle ground: minimal agentic cooperation without the overhead of large multi-agent teams. We introduce Adversarial Review (AR), a cooperative code-review protocol in which a main agent works with a reviewer and a critic. The reviewer evaluates code, while the critic audits the review through structured disagreement before the main agent edits or commits. On LiveCodeBench, AR achieves the highest pass rate among tested methods, outperforming a six-agent baseline. On SWE-PRBench, naive AR exposes a false-consensus failure mode, where agents converge on agreement without sufficient evidence, but with textual constraints it achieves the highest F1 among tested methods. On SWE-bench Verified, AR also show improvements over the baselines on repository-level coding tasks. Overall, AR demonstrates that cooperative oversight is useful when disagreement is minimal, structured, and evidence-grounded.