Inside Uber's shift from writing code to directing AI agents
Category: AI & ML
By Emily Carter
Published: 2026-07-12T06:32:00.000Z
Uber has spent the past year quietly rearranging how its engineers work, and the shape now emerging looks less like a team of coders than a team of people who direct fleets of software agents. Virtually every engineer uses AI tools, and agents now write most pull requests.
Uber has spent the past year quietly rearranging how its engineers work, and the shape now emerging looks less like a team of coders than a team of people who direct fleets of software agents. Chief technology officer Praveen Neppalli Naga recently shared internal numbers that make the shift concrete, saying that virtually every engineer at the company, some 99 per cent by his count, now uses AI tools, and that more than 70 per cent of pull requests are attributed to local or cloud-based agents. Engineers have collectively built over 2,500 agent skills covering the software development lifecycle, turning what was once individual craft into something closer to orchestration. The mechanics of the change are worth understanding. Where a developer once spent most of their day inside an editor writing code, many at Uber now kick off several agents in parallel, each handed its own task, then review and approve what comes back. The company built an internal stack to make this possible, including a gateway based on the Model Context Protocol that lets agents reach internal data, a no-code builder so non-engineers can create their own agents, and tools like Validator and Autocover that flag security issues and generate tests automatically. Uber reckons the test-authoring tool alone has saved more than 21,000 developer hours, and it deliberately points agents at the repetitive maintenance work, the migrations, upgrades and bug fixes, that developers tend to avoid. The more radical move is that Uber has stopped treating this as purely an engineering story. Over two months it ran 16 so-called agentic pods, embedding around 30 of its most AI-proficient engineers directly inside finance, legal, human resources and other functions for two weeks at a time. The results it reported are striking, with a capital-allocation analysis across 150 cities falling from 15 hours to 30 minutes, and financial pacing reports that once took two days now taking ten minutes. What surprised the team was not the speed but the fresh eyes, as engineers dropped into unfamiliar departments and spotted inefficiencies insiders had long stopped noticing. None of this comes free, and Uber has been unusually candid about the bill. Reports indicate the company burned through its entire 2026 AI coding budget within the first four months of the year, driven by heavy use of tools like Claude Code, while its chief operating officer has admitted it is getting harder to prove the spending delivers proportionate returns. For the Gulf, where Uber operates through its subsidiary Careem across the UAE, Saudi Arabia and beyond, the experiment offers a template regional firms are watching closely, since sovereign AI ambitions in Riyadh and Abu Dhabi increasingly turn on exactly this question of whether agentic productivity justifies its soaring token costs.