Summary
Will Bond and Ameya Ketkar from Uber present uReview, their in-house multi-agent automated code review system built to address a measurable bottleneck: first-time-to-review latency at Uber grew from 3 hours in 2024 to 9 hours by 2026, as PR volume and size surged across thousands of engineers working in six language-specific monorepos spread across 12 global sites.
The talk walks through uReview’s full architecture: requests flow in from GitHub, Fabricator (Uber’s legacy code review tool), and an agent loop into a central review service. Multiple specialized generators — tuned for different cost and performance profiles — produce comments that pass through post-processing steps including rating, categorization, filtering, and deduplication before engineers ever see them. A key design goal is decentralized customization: teams can define AI linters using few-shot examples, link agents to team-specific knowledge bases and past PRs, and write fully custom agents tied to Uber’s existing ownership model.
Ketkar covers the observability and evaluation journey, from early surface-level NPS surveys and Google Forms to tracking developer sentiment on individual uReview comments and capturing full agent trajectory logs. Critical lessons include the need for per-team style guides and guardrails baked into the agent (models confidently produce wrong reviews without them), and ensuring security and compliance checks run consistently across all PRs regardless of which team owns them.
📺 Source: AI Engineer · Published August 28, 2026
🏷️ Format: Hands On Build
