Skills
Three agents with structured disagreement beat a five-agent review panel on LiveCodeBench
Adversarial Review puts a reviewer and a critic around the main coding agent, where the critic audits the review through structured, evidence-grounded disagreement before any edit lands. It took the highest pass rate on LiveCodeBench while using three agents against a five-agent baseline, and the highest F1 on SWE-PRBench once an explicit disagreement prompt iteration was added. The conclusion for anyone building review pipelines is that the win comes from making disagreement minimal, structured, and evidence-grounded, not from adding agents.
↳ Follow the thread