Scaling Automated Code Review at Uber with uReview
AI Engineergo watch the original →
the gist
Uber built an internal multi-agent code review engine that processes 25,000 comments weekly by prioritizing observability, team-specific customizations, and feedback-driven prompt tuning.
The Breakthrough
Uber developed uReview, a custom multi-agent code review system, to solve the bottleneck of increasing PR review times by integrating automated, team-specific review rules directly into their existing developer workflow.
What Actually Worked
- Feedback-Driven Observability: The team moved beyond surface-level metrics like cost by tracking comment addressal rates, reply sentiment, and agent trajectory (tool calls and reasoning) to identify where agents were failing.
- Team-Specific Customizations: Instead of a centralized review policy, they enabled teams to define their own review skills, link them to internal knowledge bases, and collocate these rules within the codebase to ensure domain-specific patterns and anti-patterns were enforced.
- Intelligent Post-Processing: To manage high comment volume, the system routes requests through various generators and applies a multi-step post-processing layer to rate, categorize, filter, and deduplicate comments before surfacing them to engineers.
- Guardrails and Routing: They implemented deterministic routing to match PRs with the appropriate model and generator, while enforcing strict time-bound guardrails to prevent agents from wasting compute on low-value tasks or infinite loops.
Before / After
- Cost: Reduced by 60% compared to the naive first build.
- Accuracy: Improved by approximately 70% through iterative tuning.
- Performance: The system now generates 25,000 comments per week, with a 67% overall addressal rate and 75% addressal rate for high-severity issues.
Context
As PR volume and size grew, Uber's first-review wait time increased from three hours in 2024 to nine hours in 2026. Because existing vendors lacked support for their legacy Phabricator setup and failed to provide the necessary consistency for an agentic SDLC, Uber built uReview. The system now serves as a bridge, allowing human engineers to shift their focus from mechanical implementation details to high-level architecture and product thinking as the agentic loop matures.