Uber's median time to first review went from three hours in 2024 to nine hours now, and @wbond and Ameya Ketkar of Uber built uReview, a multi-agent code review system, to claw that back. @aiDotEngineer has their talk, "Building uReview, Uber's Multi-Agent Code Review Engine," on YouTube. It's a detailed look at running automated review across thousands of engineers, hundreds of teams, and six language-specific monorepos, with real numbers on cost, quality, and where humans still fit. - Why build in-house. Uber still runs on Fabricator while migrating to GitHub, wants agents in the inner loop to get the same review rules as humans, and needs to plug into a distributed, team-owned rule system instead of centralizing it. - How a review actually runs. Requests get routed to different generators tuned for cost and performance, then post-processed through rating, categorizing, filtering, and deduplication so engineers only see the highest-confidence comments. - Observability got deeper over time. UReview started as a single prompt judged by cost and Google Form surveys, then grew to tracking reply sentiment, addressal rate, and full agent trajectories to see why the agent did what it did. - A core lesson. The model doesn't know when it's wrong, so each team's style guide and anti-patterns have to be baked in, along with guardrails so the agent doesn't burn turns on the wrong things. - The customization stack. Single-file and multi-file reviewers, few-shot AI linters for deterministic rule checks, and fully custom agents teams can link to their own knowledge base and past PRs. - The numbers. Around 25,000 comments a week, a 67% addressal rate, roughly three-quarters of high-severity issues addressed, costs down 60%, and quality up about 70% versus their naive baseline. - Inner loop vs. outer loop. As agents write and review more code, inner-loop accuracy matters more to avoid fix-review-fix oscillation, and the outer loop doesn't disappear, it shifts toward architecture, domain expertise, and product thinking. I'm working through the published talks from AI Engineer World's Fair sharing summaries and takeaways. Follow for more!