LinkedIn's multi-agent AI code reviewer runs 79,000 reviews a week with 63.9% of suggestions accepted
As AI coding agents pushed more code through LinkedIn's 10,000+ active repositories and P90 time to first human review rose in Q4 2025, LinkedIn's Developer Productivity Automation team built its own multi-agent code reviewer instead of a single-vendor product. An orchestrator runs several model-and-harness subagents in parallel without sharing outputs, treats overlapping findings as high confidence, verifies unique ones, and layers a three-tier customization framework of org, repo and case-specific rules. It runs event-driven on a Kubernetes worker pool (12 concurrent PRs per host), starts within about 90 seconds and usually finishes in under 10 minutes, with paged on-call and staged rollouts. Every change passes a blind pairwise LLM-judge eval before shipping; a post-merge eval of 5,230 sampled comments across 1,727 PRs found 63.9% accepted. Human review is still required on every PR.
The numbers (as reported)
| Weekly reviews | 79,000+ across 40,000+ PRs in 7,500+ repositories |
|---|---|
| Repositories adopted | ~10,000 |
| Combined acceptance rate | 63.9% (sampled) |
| Task completion rate | 99.1% |
| Review latency | Starts within ~90 s, most finish under 10 minutes |
| Acceptance by category | Logic errors 80%, bug fixes 58.1%, refactoring 43.5%, security 40.6% |
Facts
- Company / subject
- Tools
- Unstated
- Codebase
- brownfield
- Work type
- Company rollout
- Stack
- GitHub, Kubernetes
- Published
- 2026-08-13
- Last checked
- 2026-11-06
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- 63.9% combined acceptance rate for automated review suggestions on sampled PRs 63.9%
- Most reviews are posted before the first human reviewer opens the PR
“Most reviews are posted before the first human reviewer opens the PR.”
Caveats
Builder-written. The models and harnesses behind the subagents are not named. This is review of code, not agent-written code; acceptance was judged by an LLM evaluator on a 7-day sample.
More stories by tool: Unstated
Related topics: Company rollouts and costBrownfield codebases