AI coding has made CI a bottleneck, so we reworked ours to keep up

Here’s the second-order effect of AI coding nobody budgets for: CI becomes the bottleneck. Linear’s test suite nearly quadrupled this year — about 2,000 new tests a week — so they reworked the pipeline instead of waiting it out: median gating jobs from 26s to 8s, per-shard setup down roughly 44%, and 87,000 runner-minutes saved a month (11.8% of total CI usage). Agents write the code. Someone still pays to validate it.