- Uber says review has become a major bottleneck as AI agents raise the volume of code changes. It is building risk scores and evidence to help.
- Uday Kiran Medisetty argued that checking agent work quickly is now a core engineering skill. Uber still has a human review every change.
When AI agents write more code, writing stops being the hard part. Judging whether a change is safe to ship takes its place. Uday Kiran Medisetty, a Distinguished Engineer at Uber, described this on The Pragmatic Engineer podcast, in an episode about Uber's AI developer tools.
What was said
Medisetty said Uber's volume of pull requests (proposed code changes) has risen year over year. Reviews, he said, have become a big bottleneck, as at any other company. Automated test jobs at least doubled as well, which pushed Uber to reinforce the systems that run them.
Uber's answer is to move checking earlier and give reviewers evidence. A single command lets an agent run style, coverage and build checks before it opens a change. Agents can also deploy their work to a test environment and try it end to end.
A tool called Code Inbox then shows the reviewer what already happened: whether the AI review passed, a test screenshot, a video and build results. Medisetty said Uber is adding a risk score and a blast radius, meaning how much a change could affect. These help engineers decide where to focus.
He tied this to the job itself. His own work now needs context-setting for agents, clear specs, and juggling many tasks at once. He also named fast verification of agent output as a skill that has become much more important. Without it, he said, regressions reach production.
Why it matters
Our reading: the scarce resource is no longer people who can produce code. It is people who can check code at the pace agents produce it. Uber's approach treats review as something to design and fund, with test environments, automated checks and clear signals.
For leaders buying or rolling out coding agents, the lesson is to budget for the checking side too. A tool that doubles output without growing review capacity may only move the queue. It also changes what to teach: judging evidence is a skill, not a side effect of writing code.
The other side
Uber is cautious about how far this goes. Medisetty said "the cost of an incident for uber is like really high", so an engineer still reviews every change. Skipping review for low-risk changes, which Uber calls a green lane, is something it might enable later. That depends on trusting its risk scoring, which it wants to test against years of past data. The scores are still being added.
A speaker on the show noted that engineers cannot review every change with the same rigor as before. Uneven attention is built into the design. It works only if the risk signals are right.
The episode also suggested Uber has an advantage. Its host observed that firms gaining most from AI had developer platform teams before AI. Many companies lack that base. The conversation did not address how newer engineers build judgment if they write less code themselves.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- The Pragmatic Engineer: Uber’s AI infra, from the inside (2026-10-14)
