LinkedIn describes a multi-agent AI code review system deployed across roughly 10,000 repositories, handling 79,000+ weekly review comments with a 63.9% acceptance rate. The architecture runs multiple model-and-harness subagents in parallel per PR, cross-validates and deduplicates their findings (treating overlap as a confidence signal), and layers a three-tier customization framework encoding org- and repo-specific conventions. Reviews typically post before human reviewers even open the PR, with a speed SLA targeting under 10 minutes end-to-end. Acceptance is measured post-merge by checking whether suggestions were actually incorporated into merged code, with correctness-related suggestions (bug fixes, logic errors, concurrency bugs) seeing far higher acceptance than refactoring or security suggestions. Future work includes a visual risk-heatmap review surface and pipelines that turn production incident postmortems into new review rules.
Table of contents
Why Not Use an Off-the-shelf AI Reviewer?Multi-Agent Architecture: Why One Model Isn’t EnoughQuestions this post answers
How does LinkedIn measure whether AI code review suggestions are actually useful, not just upvoted?
LinkedIn built a post-merge acceptance-rate pipeline that compares AI reviewer suggestions against the final merged codebase rather than relying on thumbs-up reactions. Across 1,727 PRs and 5,230 sampled comments over a 7-day window, an LLM-based evaluator classified each suggestion as accepted, partially accepted, not accepted, or not actionable, with 90.1% of judgments rated high-confidence, yielding an overall 63.9% acceptance rate. daily.dev surfaces engineering writeups like this for teams weighing how to measure AI code review impact.
Why did LinkedIn build its own multi-agent AI code reviewer instead of using an off-the-shelf vendor tool?
Single-model reviewers carry fixed training biases that cause consistent blind spots, their rule-file context injection can't compose organization-wide and repo-specific policies across thousands of repositories, and they are consumed as products rather than operated as infrastructure, preventing canarying, failover, or fleet-wide observability. LinkedIn's multi-agent system runs several independent model-and-harness subagents per PR and cross-validates their output to close these gaps. developers evaluating build-vs-buy for AI code review can compare architectures like this on daily.dev.
Which types of AI code review suggestions get accepted most often by developers?
Correctness-oriented suggestions see the highest acceptance: concurrency bug flags reach 100% acceptance, logic error suggestions reach 80%, and bug fixes overall sit at 58.1%, the highest-volume category. Refactoring suggestions land at 43.5% and security suggestions at 40.6%, reflecting that these categories are more subjective and context-dependent rather than indicating lower suggestion quality. teams tuning AI reviewer prompts can track category-level acceptance benchmarks like these on daily.dev.
10 Impressions1 Comment