Preetam Dwivedi, staff engineer on Uber's developer platform, told AI Engineer that AI-written code broke his merge math. He said SubmitQueue took 90 requests an hour in January 2026 and 130 by March, a 44% jump tied to AI diffs, while P95 land time stretched from 30 minutes to three to four hours.
The trunk stayed green. Developers stalled.
Dwivedi described more than 5,000 developers landing over 36,000 commits and 27,000 PRs a week across six monorepos, with Go as the scaling pain. He said three older tricks carried the human era, with target analysis to run disjoint builds in parallel, probabilistic speculation to run likely builds first, and bypass to let small changes jump large ones.
Manjari Akella, senior engineer alongside Dwivedi, said the team answered AI volume with admission gating and deferrals. She said gating sorts small changes from expensive ones and turns expensive ones away at peak, which lifted business-hours throughput from 130 to about 160 lands an hour, while deferrals persist rejects in a second queue and replay them when capacity frees.
She said conflict relaxation bought the real gain by scoring overlaps from zero to one with Jaccard similarity and cutting weak links to force parallel lanes, taking peak to almost 250 and sometimes 300 lands an hour with P95 back at 30 minutes. She conceded the cost, she said, since skipped combinations can land untested and break main, so StowPipe validates post-land and holds a verified main pointer until builds pass.
The queue pain is familiar outside code. Drivers who work Uber and Lyft argue the same throughput math at airports, with one driver account saying hour-plus airport waits rarely pay so he skips them after drop-off. Uber chose the opposite for code, taking more volume and cleaning up after.
Dwivedi still framed a green trunk as the goal. Akella treated a broken trunk as the price and left recovery manual, with on-call paged to fix or revert. Neither said how often AI diffs cause those breaks, or who should own them when 5,000 humans share blame with models. Who gets paged when the bot that wrote the code cannot revert it?
