Thirty-one percent of Uber's code is now written by AI. Ninety-two percent of the company's developers use AI agents at least monthly. And 11% of pull requests are opened not by humans, but by automated agents.
Those numbers come from an inside look at how Uber has rebuilt its development workflow around AI tooling. The company's stated goal is to become a "GenAI-powered" company, with AI handling what leadership calls the "boring" work: upgrades, migrations, trivial bug fixes, and boilerplate tests.
The Internal Tool Stack
Uber did not just hand developers ChatGPT and call it a day. The company built a full internal platform:
- MCP Gateway connects internal APIs (Thrift, Protobuf, HTTP) to AI models through the Model Context Protocol, handling auth, logging, and telemetry centrally. This means agents can interact with Uber's internal services the same way a developer would.
- Uber Agent Builder is a no-code platform for creating agents that access internal data and delegate tasks to other agents. It includes visualization, debugging, tracing, and version control.
- AIFX CLI handles agent provisioning, MCP server discovery, and background task execution across the entire engineering org.
- Autocover generates over 5,000 unit tests per month automatically.
- Minion runs background agents with full monorepo access for large-scale tasks.
- uReview adds AI-generated code review comments, and Code Inbox handles smart PR routing.
How Daily Work Changed
The biggest shift is not any single tool but the workflow itself. Uber's engineers have moved from single-threaded IDE work to orchestrating multiple agents running in parallel. A developer kicks off one agent, starts another task, checks back when results arrive, then chains the next agent. It is closer to project management than traditional coding.
Director Anshu Chada put it bluntly: "When we push boring stuff to AI...engineers focus on product features in ways we didn't think possible."
The Cost Problem Nobody Talks About
Here is the number that should get attention alongside the productivity stats: Uber's AI-related costs have increased 6x since 2024. Token cost optimization is now an emerging priority internally.
This is the reality check hiding behind every "X% of our code is AI-generated" headline. AI coding tools produce results, but they burn through tokens at scale, and those costs compound fast across thousands of engineers. Uber also found that adoption was slower than expected despite a forward-thinking engineering culture. Top-down mandates were less effective than peer-shared wins, suggesting that AI tool adoption spreads more through hallway conversations than executive memos.
For engineering teams watching Uber's numbers and thinking about their own AI strategy, the takeaway is not just "adopt AI tools" but "budget for the infrastructure and cost management that comes with them." The productivity gains are real. So is the bill.