LinkedIn's multi-agent AI code review: 63.9% of sampled suggestions accepted by developers
LinkedIn built an AI code reviewer that runs several AI models in parallel on each pull request and applies team-specific rules. In August 2026 it reported a 63.9% combined acceptance rate on a one-week sample, with more than 79,000 reviews a week.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
63.9%
Combined acceptance rate of the AI reviewer's comments on sampled pull requests (1,727 PRs, 5,230 sampled comments, a 7-day window)1
58.1%
Acceptance rate for bug-fix suggestions, the highest-volume category1
80%
Acceptance rate for logic-error suggestions1
43.5%
Acceptance rate for refactoring suggestions1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Code review is a bottleneck for software delivery. LinkedIn has more than 10,000 active code repositories and tens of thousands of pull requests (proposed code changes) every week. AI coding tools were adding more code per developer, and the wait for a first human review was rising. Generic AI reviewers produced comments that missed each team's own conventions, and noisy comments train developers to ignore the tool.
What they built
A multi-agent reviewer. An orchestrator sends each pull request to several review sub-agents at the same time. Each sub-agent is a different pairing of an AI model and its surrounding tooling. The orchestrator collects their findings, removes duplicates, checks them and posts one review. LinkedIn compared this with the hosted review products from model vendors and chose to build its own. It gave three limits: each carries its model's blind spots, one rule file cannot express company-wide and per-repository standards, and a bought product cannot be tested on a few repositories first or monitored across the fleet.
How they did it
Findings that several sub-agents raise independently are treated as high confidence. A finding raised by only one is checked before it is posted. A three-tier rules framework lets the company, groups of teams and single repositories each add review guidance, and combines them for each change. The system runs on GitHub events and a pool of workers on Kubernetes. Each worker machine handles as many as 12 pull requests in parallel. LinkedIn runs it with paged on-call, monitoring of speed and acceptance, staged rollouts and a fallback when a model provider degrades.
Quality is checked in two ways. Before any change to a prompt or model ships, it must beat or tie the current version in a blind, head-to-head comparison scored by an AI judge. After release, an automated check compares each suggestion with the code that was finally merged, to see whether developers acted on it.
Results
LinkedIn reports a 63.9% combined acceptance rate on a sample of 1,727 pull requests and 5,230 comments over seven days. By category, logic errors reached 80% and bug fixes 58.1%, while refactoring (43.5%) and security (40.6%) were lower. LinkedIn reads the lower figures as a sign these areas are more subjective. It reports that in a typical week the system completes 79,000+ reviews across 40,000+ pull requests in 7,500+ repositories, with a 99.1% task completion rate. It also reports that most reviews finish in under 10 minutes, and that the AI review usually lands ahead of any human reviewer opening the change. Acceptance varied from 0% to 100% across individual repositories. LinkedIn says many engineers now wait for the AI review before asking a colleague, and that human review remains required on every change.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Combined acceptance rate of the AI reviewer's comments on sampled pull requests (1,727 PRs, 5,230 sampled comments, a 7-day window) | 63.9%1 | |
| Acceptance rate for bug-fix suggestions, the highest-volume category | 58.1%1 | |
| Acceptance rate for logic-error suggestions | 80%1 | |
| Acceptance rate for refactoring suggestions | 43.5%1 | |
| Acceptance rate for security suggestions | 40.6%1 | |
| Task completion rate in a typical week | 99.1%1 | |
| Time for most reviews to complete end to end | under 10 minutes1 |
Lessons for your project
Measure whether people act on AI output, not whether they like it. LinkedIn rejected thumbs-up reactions as a vanity metric and checked the merged code to see which suggestions were adopted.
Do not rely on one model. LinkedIn runs several model and tool pairings on each change and trusts findings that more than one of them raise independently.
Let each team encode its own standards. A payments service and a front-end experiment need different review rules, and acceptance varied from 0% to 100% across repositories.
Speed decides trust. The reviewer is designed to post before the human reviewer opens the change, because a late AI review is treated as noise.
Operate the system like any critical service: on-call staff, monitoring, staged rollouts of prompt and model changes, and a fallback when a model provider slows down.
Caveats
- All figures are LinkedIn's own, from one engineering blog post (InfoQ summarises the same post and does not add independent data), and cover one 7-day sample. Acceptance was judged by an AI-based evaluator comparing suggestions to the merged code; the post says 90.1% of judgments were rated high confidence.
- The post does not say which models or vendors the reviewer uses. It describes several different model-and-tool pairings. The vendor is shown as undisclosed.
- Acceptance rate is not an accuracy rate. LinkedIn says the roughly 37% not adopted are not all AI mistakes, and that every change still gets human review.
- No source gives a go-live date. The deployed year (2026) is taken from the date of LinkedIn's post, and the system may have gone live earlier: the post describes the review backlog as growing in Q4 2025 and refers to an earlier company-wide announcement.
- The post states the system was adopted by about 10,000 repositories, while its weekly figure covers 7,500+ repositories. We show the weekly figure as scale.
- No cost or team-size figure is given for the build. The post thanks 14 named team members and a manager in its acknowledgements, without saying how much of their time it took.
Sources
- High-signal AI code review that adapts to your codebase at scale
LinkedIn Engineering Blog · 2026-08-13 · Company blog · linkedin.com
- AI Code Review at Scale: LinkedIn's Multi-Agent Approach
InfoQ · 2026-08-22 · News · infoq.com
Summarized in our own words from the sources above. Spotted an error? See corrections.