Adversarial Review: Small Agent Teams Beat Larger Ones in Code Review

rohanpaul_ai · x · 2026-08-24

A paper proposes the 'Adversarial Review' architecture: code is frozen while a reviewer writes a critique and a critic audits it, with only the settled review returned for edits. Research shows that simply adding agents leads to false consensus; the key is a rule requiring objections to be based on code evidence, not hunches. On LiveCodeBench, a 3-agent version reached 87% (beating a 5-agent version's 82%). In PR review, enforcing the code evidence rule boosted F1 from 0.457 to 0.533.

Related event: Structured Adversarial Multi-Agent Code Review Beats Scale(2 posts)→

Original post →

More from coding & agent

coding & agent channel →