More review agents can mean less real disagreement
A new paper finds naive multi-agent review collapses into false consensus, and that a smaller setup with an explicit critic does better.
Source: Adversarial Review: Multi-Agent Code Review with Explicit Dissent — arXiv 2608.18167
A common instinct in agentic code review is to add more reviewers. This paper argues that without a role built for dissent, those reviewers often agree without evidence.
The setup is deliberately small: a coder, a reviewer, and a critic. The critic does not rewrite the reviewer’s comments. Its job is to challenge the proposed change with structured, evidence-grounded disagreement so a diff does not get accepted on an easy consensus. On their benchmarks, three agents with that explicit dissent beat a five-agent team. A single prompt change that forces disagreement also lifts review F1.
Why it matters
If your review loop cannot name what it is disagreeing about and point at evidence, it is optimizing for social smoothness. For teams building AI review into a software factory, the useful primitive is not “more agents.” It is disagreement as a first-class protocol.