Uber engineers' Claude Code use nearly doubled in three months as IDE tools plateaued

The 84% figure rests on a two-part definition. A developer counts as an agentic coding user by using CLI-based agents, or by making more agentic requests than tab-completion requests in the IDE ^[2]. The second test is a threshold on habit inside an editor, and a developer can qualify while still using tab-completion ^[2]. The same piece lists 92% of Uber developers as using agents monthly ^[8], eight points higher ^[19]. The two figures use different definitions.

The 65-72% share of AI-generated code covers IDE-based tools only ^[3]. For command line tools such as Claude Code, the newsletter says the share is, naturally, 100% ^[3]. Claude Code use rose 31 points between December and February ^[18], so any blended share would climb as engineers move to CLI agents, even if nobody changes habits inside a given tool. We would read the AI-generated share as a measure of tool mix as much as of output.

Claude Code usage went from 32% in December to 63% in February ^[4], a rise of about 1.97 times ^[18]. Cursor and IntelliJ plateaued over the same period ^[5]. The newsletter reports that top-down mandates were less efficient than engineers sharing their wins with peers ^[15]. A standard written around December's mix would have been out of date by February ^[18].

Cost sits alongside the adoption numbers. Uber's spending on AI has risen 6x since 2024, and bringing down token costs is becoming a bigger priority for the company ^[10]. The newsletter says engineers naturally gravitate toward kicking off new agents, and that this starts to create resource and cost challenges ^[11]. Uber did not disclose dollar amounts for the increase.

All of these figures are Uber's own, relayed by a newsletter whose author worked at Uber until 2020 ^[17]. Uber is also a cautious stand-in for other firms. It employs almost 3,000 people in its tech function ^[6], and its engineers were candid about how much work it took to build the internal AI stack ^[16]. The newsletter says adoption was slower than expected even at a forward-thinking company like Uber ^[14].

We think the direction in Uber's data carries over to other firms. The levels depend on an internal platform that took Uber substantial work to build ^[16]. The decision this quarter is whether to standardise on a tool or leave room for the mix to move, and the December-to-February data favours leaving room ^[18]. Any productivity case built on these usage figures has to be set against a 6x cost rise since 2024 ^[10].