Great paper on multi-agent systems for code review.

It's challenging to know how many coding agents to use to address a problem.

The default fix for weak agentic code review is more agents. In turns out that scaling agents to a large number gives diminishing returns on repository-level tasks.

This new work tries structured conflict instead. Adversarial Review runs three agents. A main coding agent writes, a reviewer evaluates, and a critic audits the review before any edit are done.

On LiveCodeBench it beats a five-agent baseline while using three agents.

On SWE-PRBench the naive version exposed a failure mode. The agents converged on agreement without enough evidence behind it. Making disagreement an explicit instruction recovered the highest F1 among tested methods.

They also find that cooperative review works when the disagreement is minimal, structured, and grounded in evidence.

Paper: https://arxiv.org/abs/2608.18167

Track more trending AI papers in our academy: https://academy.dair.ai/