Agentic AI
Three-Agent Code Review System Outperforms Larger Teams by Embedding Structured Disagreement
Researchers show that effective AI-driven code review doesn't require teams of many agents—just a lean setup where agents actively disagree in evidence-grounded ways.
A new research framework called Adversarial Review (AR) challenges the assumption that multi-agent AI code review needs to be complex to be effective. Instead of scaling up agent teams, the approach achieves better results with only three specialized agents working in a tightly structured protocol.
The Problem with Traditional Multi-Agent Setups#
Earlier attempts at using large language models for collaborative code review relied on teams with separate roles—each agent assigned a specific function. However, as researchers added more agents, the gains diminished quickly when tackling real-world repository-level programming tasks. More recently, some systems have swung toward the opposite extreme, treating agents as simple passive tools without meaningful interaction. Both paths have drawbacks: one gets bogged down in coordination overhead, the other loses the benefits that come from agents pushing back on each other's ideas.
How Adversarial Review Works#
AR introduces a middle path through minimal but deliberate cooperation. The system pairs a main coding agent (which writes and edits code) with two specialized partners: a reviewer who examines the code and a critic who directly audits the reviewer's own assessment. The critic doesn't simply agree or disagree—it operates through structured disagreement, forcing explicit reasoning about why positions differ and grounding all feedback in concrete evidence before the main agent makes any edits.
Performance Across Benchmarks#
Testing on three major benchmarks revealed consistent wins:
- LiveCodeBench: AR achieved the highest pass rate among all tested approaches while using three agents instead of five in competing methods.
- SWE-PRBench: The initial naive version exposed a failure mode where all agents would converge on agreement without adequate justification. A single prompt adjustment that made disagreement mandatory solved this, yielding the best F1 score among all approaches.
- SWE-bench Verified: The framework also demonstrated measurable improvements over baselines on full repository-level coding tasks.
