Uber's uReview Rebuilds Automated Code Review

Uber (NYSE:UBER) saw first time to review triple from 3 hours in 2024 to 9 hours in 2026, forcing automated code review from experiment to bottleneck fix.

Will Bond and Ameya Ketkar detailed the system at AI Engineer, where they showed how uReview now serves thousands of engineers across hundreds of teams and 12 sites.

Why Uber built instead of bought

Uber still runs Fabricator but is migrating to GitHub, and off the shelf reviewers do not support Fabricator.

The company also needs parity between human and agent loops so its agentic SDLC gets the same rules in both places.

Centralized rule management fails at Uber scale, so uReview plugs into the existing team ownership graph and routes by risk and complexity.

How uReview works under the hood

Requests from GitHub, Fabricator and the agent loop hit a central uReview service that dispatches to multiple generators tuned for different cost and performance tradeoffs.

It can also call third party reviewers for comparison, then post processes with rating, categorizing, filtering and deduplication to avoid flooding engineers with low confidence nits.

That pipeline now emits about 25,000 comments per week, a volume that makes filtering non optional.

From prompt to production: observability wins

Early uReview was a single prompt for per file logic checks with a simple dispatcher and surface metrics like cost and NPS via Google Forms and Slack.

The team added sentiment analysis on developer replies, then tracked address rate and full agent trajectories including tool calls and reasoning to tune latency and quality.

The model always sounds certain, so Uber added team specific style guides and guardrails to stop wasted turns, cutting costs 60% and lifting accuracy 70% versus the naive baseline.

Customization at scale is the hard part

Uber runs six language specific monorepos, so uReview layers single file logic bug finders, multi file deep reviewers with monorepo anti patterns, few shot AI linters for mechanical issues, and fully custom agents linked to knowledge bases and past PRs.

Customizations live next to code and ride the ownership graph, with deterministic routing that decides which team gets which generator and model.

Writing a skill is easy with Claude drafting from past reviews, but running it with consistent quality and low cost at scale required iterative tuning from both the platform team and the owning teams.

Does it actually help

About 10% of the 25,000 weekly comments get explicit feedback and only 4% of PRs draw negative feedback.

The address rate sits at 67% overall and roughly three quarters for high severity issues, which Uber reads as signal that comments are actionable.

Those numbers improved only after exposing address rate, sentiment and trajectory data back to teams so rule authors could fix or retire noisy checks.

Why this matters for the inner and outer loop

Uber still requires human approval today, but Bond flagged a near term path where a percentage of code lands with automatic approval as agents author more code.

Accuracy demands rise in that inner loop because bad comments cause cavitation where an agent fixes, gets re reviewed and then reverts, while agents happily fix 100 nits that would exhaust a human.

Instead of killing the outer loop, Uber wants to expand it so engineers review architecture and product tradeoffs rather than API compatibility and performance nits.

How Uber stacks up

StartupHub.ai data shows Uber at 76/100, just behind DoorDash (NASDAQ:DASH) at 78/100 and ahead of Lyft (NASDAQ:LYFT) at 72/100, Grab (NASDAQ:GRAB) at 66/100, Gett at 54/100 and Careem at 55/100 among tracked mobility peers.

That middle to top positioning reflects a company with enough scale to justify custom tooling where smaller peers can buy Graphite, CodeRabbit or GitHub Copilot code review off the shelf.

The bet is that owning the review layer lets Uber enforce security and compliance consistently while distributing taste, a combination vendors still struggle to offer across legacy systems like Fabricator.

If uReview keeps high severity address rates near 75% while holding cost down, Uber's next bottleneck will not be review latency but governance of who gets to merge without a human.