At LinkedIn's scale, relying solely on human reviewers or simply putting an off-the-shelf AI reviewer in front of GitHub is not an effective way to manage PRs. To address this, LinkedIn engineers built a multi-agent AI code review platform that understands the organization’s coding context, treats code review as production infrastructure, and minimizes hallucinations and low-signal feedback.
LinkedIn's goal with its code review platform is generating reviews that developers find worth acting on, maximizing signal-to-noise, and accounting for the codebase's "standards, conventions, and tribal knowledge that generic AI models consistently miss".
Generating AI review comments at scale is trivial. 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.
Specifically, relying on an off-the-shelf AI reviewer brings three structural limitations to the fore: blind spots caused by using a single model, which can lead the AI reviewer to miss the same class of bugs and flag the same low-signal issues; insufficient customization, which makes it difficult to simultaneously encode organization-wide policies, repository-specific conventions, and targeted guidance for high-risk or high-impact scenarios; lack of operational control, which constrains the ability to control, evaluate, and monitor the reviewer as part of the engineering infrastructure.
Based on this analysis, LinkedIn's platform addresses each of these limitations through multiple independent AI reviewers that use distinct models and reasoning approaches; deep, composable customization spanning organization-wide policies, repository-level conventions, and context-specific rules; and a Kubernetes-based architecture that supports an event-driven pipeline with durable queues and horizontally scaled workers, enabling monitoring of latency, acceptance and completion rates, as well as provider failures.
Using multiple independent AI reviewers enables cross-validation, increasing confidence in the findings. For example, when multiple agents independently identify the same issue, LinkedIn treats that convergence as strong evidence. Unique findings, however, aren’t automatically discarded but verified separately. Finally, cosmetic, already-fixed, irrelevant, or repository-inconsistent suggestions get filtered before posting.
To measure how often developers actually implement AI-generated suggestions, LinkedIn built an automated acceptance-rate evaluation pipeline to compare all suggestions against the resulting merged codebase. The evaluation covered 5,230 sampled review comments across 1,727 PRs and found that 90.1% could be evaluated with high confidence based on the merged code. Overall, 63.9% of suggestions were accepted, with significant variation by category: 80% of logic errors, 58,1% of bug fixes, 43,5% of refactoring changes, 40,6% of security-related fixes, and 100% of concurrency bugs were accepted.
Other companies have also tackled the problem of code review at scale, but with different approaches. For example, Cloudflare built an orchestration system around the open-source coding agent OpenCode, striking a different balance across its requirements and constraints. Similarly, Databricks released several components, including the Unity AI Gateway for centralized AI management and Omnigent for developer tooling, addressing what it describes as the "exponential growth of AI coding costs".
There is much more to LinkedIn's multi-agent review platform than can be covered here. For the full technical details and implementation insights, be sure to read the original article.