LinkedIn's Multi-Agent AI Code Review System Posts 79,000 Reviews a Week, With Developers Accepting 64% of Suggestions

Overview

LinkedIn has detailed a multi-agent AI code review platform that it says completes more than 79,000 reviews a week across over 40,000 pull requests spanning more than 7,500 repositories, according to a LinkedIn Engineering blog post authored by engineer Min Chen. The company reports a 99.1% task completion rate for the system and, in a sampled evaluation, a 63.9% rate at which developers accepted the AI-generated suggestions, according to the same post and InfoQ’s coverage of it.

What We Know

• LinkedIn operates more than 10,000 active repositories with “tens of thousands of pull requests opened every week,” and the company says the volume is growing as AI coding agents increase the amount of code shipped per developer, according to LinkedIn’s post.
• In a typical week, the platform completes “79,000+ reviews across 40,000+ PRs spanning 7,500+ repositories,” at a “99.1% task completion rate,” according to LinkedIn’s post.
• LinkedIn evaluated existing hosted AI code-review products before building its own, and found three structural limitations: blind spots from relying on a single model, which can cause the reviewer to miss the same class of bugs while over-flagging low-signal issues; insufficient customization to encode organization-wide policy alongside repository-specific conventions; and a lack of operational control over monitoring, evaluation, and staged rollouts, according to InfoQ.
• The team wrote that “generating AI review comments at scale is trivial,” and that “the hard part is everything that comes after: making them factually grounded in the diff rather than hallucinated; high-signal rather than noisy; specific to the conventions of this codebase rather than generic best practices; and arriving before the human reviewer, not after,” according to LinkedIn’s post.
• The system runs multiple independent AI reviewers using distinct models and reasoning approaches, then deduplicates their output; when several agents independently flag the same issue, LinkedIn treats that convergence as stronger evidence rather than discarding unique, single-agent findings outright, according to InfoQ.
• The infrastructure is built on a Kubernetes-based, event-driven pipeline with durable queues and horizontally scaled workers. PR events are queued in under 5 seconds, a review starts within roughly 90 seconds, each worker host can review up to 12 pull requests concurrently, and most reviews complete end-to-end in under 10 minutes, according to LinkedIn’s post.
• To measure real-world impact, LinkedIn built an automated acceptance-rate pipeline that compares AI suggestions against the code developers actually merged. Sampling 5,230 review comments across 1,727 pull requests, the pipeline could evaluate 90.1% with high confidence. Overall acceptance was 63.9%, breaking down to 80% for logic errors, 100% for concurrency bugs, 58.1% for bug fixes, 43.5% for refactoring suggestions, and 40.6% for security-related fixes, according to LinkedIn’s post and InfoQ.
• LinkedIn is not alone in building custom infrastructure for AI-assisted code review at scale: Cloudflare built an orchestration system around the open-source coding agent OpenCode, while Databricks has released components including a centralized AI-management gateway and a developer-tooling component it says address the “exponential growth of AI coding costs,” according to InfoQ.

What We Don’t Know

• Neither LinkedIn nor InfoQ disclosed the specific AI models used by the independent reviewer agents, or the total engineering headcount behind the platform.
• It is not clear how LinkedIn’s 63.9% acceptance rate compares directly to acceptance rates at Cloudflare or Databricks, since neither company has published a directly comparable metric alongside LinkedIn’s.

Analysis

The acceptance-rate breakdown LinkedIn published is notable less for the headline 63.9% figure than for its spread: suggestions the platform is most confident are objectively verifiable, such as logic errors and concurrency bugs, were accepted at 80% and 100% respectively, while more subjective categories such as refactoring and security fixes saw acceptance drop to 43.5% and 40.6%. That gap illustrates a pattern the industry has been grappling with more broadly as AI coding agents generate a growing share of pull requests: automated reviewers are more reliable at catching mechanically verifiable defects than at making judgment calls a human reviewer would weigh against broader codebase context and organizational risk tolerance.