AI Can Write the PR. Your CI Still Has to Prove It.
Maestro Brief · Published by Maestro Mojo
2026-09-22
Maestro’s take
The agent finished the patch. Great. It has not shipped a thing until the checks pass. Linear says its test suite nearly quadrupled this year, yet it cut pull-request CI wait from more than six minutes to just over five and roughly halved runner time per test. Those are Linear’s own measurements, not a promise for your repo. The useful lesson: once agents make code faster, verification becomes the queue.
TL;DR
Linear found four levers: faster runners and tools, shorter jobs that block everything behind them, less repeated setup, and better-balanced tests. The clever bit is not a faster model. It is removing wasted waiting without dropping checks.
Why Maestro users care
Several agents can finish changes at once. If every patch waits on the same slow gate, adding another coding agent may simply lengthen the line. Linear says a change-detection job fell from 26 to 8 seconds at the median after it stopped fetching more of the repository than it needed. It also grouped seven tiny checks into two jobs, saving an estimated 87,000 runner-minutes a month at its scale. These figures are self-reported by Linear, not independently verified.
Do this, not that
- Do not buy more agent speed before measuring the path from PR to green. Inspect the CI job graph, the first blocking job, the slowest test shard, and runner minutes per PR. GitHub documents both its workflow graph and job execution time.
- Do not double the number of test shards just because they are slow. Each shard pays setup again. First trim checkout, dependency installs, and repeated database setup; then measure whether more shards actually help.
- Do not turn off isolation or required tests for a pretty timing chart. Linear shared test module state only for explicitly eligible files, with teardown; it left troublesome tests isolated.
One thing to try
Take your last ten PRs. Find the job that makes every test wait. Shorten that gate or move non-blocking work off the critical path. Compare time-to-green and total runner minutes before and after. Keep the checks.
Original sources
- Linear engineering, “AI coding has made CI a bottleneck, so we reworked ours to keep up” (Sep 21, 2026)
- GitHub Actions documentation, “Monitor workflows”
- GitHub Actions documentation, “Viewing job execution time”
Linear supplied every speed and cost figure in this story. GitHub’s docs support only the steps for inspecting your own workflow and job durations. Maestro’s interpretation and checklist are analysis, not our own benchmark. AI-assisted; reviewed by an independent AI editor.