LinkedIn has built a multi-agent code review platform that checks pull requests with several independent reviewers rather than relying on one general-purpose model. The system combines organization-wide policies, repository conventions, and rules for specific high-risk scenarios before deciding which comments developers should see.
Multiple agents use different models and reasoning approaches. When several reviewers identify the same issue independently, the platform treats that agreement as stronger evidence. Findings raised by only one agent are verified separately, while cosmetic, irrelevant, already-fixed, or repository-inconsistent suggestions are filtered out.
LinkedIn runs the service as production infrastructure on Kubernetes, using an event-driven pipeline, durable queues, and horizontally scaled workers. That setup lets engineers monitor latency, completion and acceptance rates, and failures from model providers instead of treating AI review as a simple GitHub add-on.
To estimate usefulness, LinkedIn compared suggestions with the code that ultimately reached merged branches. Of 5,230 sampled comments across 1,727 pull requests, 90.1% could be evaluated with high confidence, and 63.9% were accepted overall. Acceptance varied substantially by category, including 80% for logic errors and 40.6% for security fixes. The figures describe LinkedIn's sampled deployment rather than a general benchmark for AI code review.