How LinkedIn Scales AI Code Review with Multiple Agents

Introduction

At LinkedIn’s scale, the difficult part of AI-assisted code review is not producing a large number of comments. It is producing comments that developers consider credible, relevant, and worth acting on. The material describes three structural weaknesses in simply adopting an off-the-shelf reviewer: a single model can share blind spots and repeat low-value findings; generic tools often lack the ability to encode organizational and repository-specific knowledge; and they provide limited operational control over quality, latency, and failures.

LinkedIn therefore treats AI review as production infrastructure rather than as a thin layer placed in front of GitHub. Its multi-agent platform is designed to increase the signal-to-noise ratio and generate useful findings before human reviewers complete their work.

How the platform works

• Independent reviewers provide diversity. Multiple agents inspect the same change using different models or reasoning strategies. When several agents independently identify the same issue, the convergence is treated as evidence that the finding deserves greater confidence. A finding raised by only one agent is not automatically discarded; it is sent through separate validation instead.
• Local engineering knowledge becomes part of review. The platform supports composable customization at several levels: organization-wide policies, repository conventions, and rules tied to a particular context or high-risk scenario. This lets the system look beyond generic best practices and account for the practices teams have developed over time.
• Low-value feedback is filtered before publication. Cosmetic suggestions, findings already fixed in the change, irrelevant comments, and advice inconsistent with repository conventions are removed before developers see them. The goal is not to maximize model output, but to protect the review workflow from noise.
• The pipeline is operated like infrastructure. A Kubernetes-based, event-driven architecture uses persistent queues and horizontally scalable workers. This allows the team to monitor latency, acceptance rate, completion rate, and provider failures, turning the reviewer into a measurable service rather than an opaque model endpoint.

Measuring adoption instead of output volume

LinkedIn also created an automated evaluation pipeline that compares AI suggestions with the final merged code. Across a sample of 5,230 comments from 1,727 pull requests, 90.1% could be evaluated with high confidence using the merged code, and 63.9% of suggestions were accepted overall. Results varied considerably by category: logic errors reached 80%, bug fixes 58.1%, refactoring changes 43.5%, security-related fixes 40.6%, and concurrency defects 100%.

These figures do not imply that every comment has equal reliability. They do show why post-merge behavior is a more meaningful evaluation target than the number of comments generated. A review system should be judged by whether it leads to useful code changes, not by how active its language model appears.

Why it matters

LinkedIn’s approach suggests that enterprise AI coding tools are moving from model integration toward system design. Multiple agents cannot eliminate hallucinations on their own, and they introduce additional costs in orchestration, validation, and resource management. They can, however, separate review capabilities and create opportunities for independent cross-checking.

Other companies are exploring different trade-offs. Cloudflare has built an orchestration system around the open-source coding agent OpenCode, while Databricks has introduced components such as Unity AI Gateway and Omnigent in response to the rising cost of AI-assisted development. There is no single blueprint yet. The more transferable lesson is to define what counts as a valuable review, then build a feedback loop around acceptance, false positives, latency, and service reliability.

As AI-generated code becomes more common, the practical differentiator may be less about which model produces the most comments and more about which platform can consistently turn those comments into developer trust and better merged code.

Source: InfoQ Chinese