At Uber, over 70 percent of code is now agent work, and the software factory's cost report is out
Uber engineering published the numbers: more than 70 percent of pull requests are attributed to agents, 3,600 agent skills run 30,000 times a day. Usage grew 7x while total AI spend has held steady since April.
The most concrete numbers yet on how far agents have taken over enterprise software development come from Uber. The company's engineering team published a progress report on its "Software Factory" vision, and the picture is striking: more than 70 percent of pull requests at Uber are now attributed to local or cloud agents. Engineers have built over 3,600 agent skills across the software development life cycle, and those skills are executed more than 30,000 times a day.
More notable still, a growing share of sessions is not initiated by humans at all. Managed agents handle code review on their own, self-heal continuous integration failures, complete end-to-end pull requests with visual validation, triage on-call alerts and debug incoming bugs, with humans staying in the loop at review and escalation points.
Growth and cost at the same time
From February to mid August, weekly active users across all employees (engineers and non-engineers alike) grew 7x and weekly agent requests grew 9.4x. Total AI spend has nonetheless held relatively steady since April. The team also isolated how: holding one model fixed from February to July, cost per 1,000 model requests is down almost 34 percent from its peak, and cost per session is down 52 percent from its June peak.
Context: this week we published the SpaceXAI engineer's report on managing 2,000 pull requests a month with five bots, and our explainer on what a harness is. Uber's report is the same story at company scale: the question is no longer whether to use agents, but how to run thousands of them without blowing up the bill.
