How Linear Reworked CI to Keep Pace with AI Coding

Summary

Linear says AI coding agents have accelerated code shipping faster than its continuous integration system could validate changes, making CI both a developer bottleneck and a source of rising infrastructure costs. Despite its test suite nearly quadrupling since the start of the year, the company reduced pull-request wait time from more than six minutes to just over five and roughly halved runner time per test. It moved workloads from GitHub Actions to third-party runners, making jobs 34% faster on average, and adopted the native TypeScript compiler tsgo, cutting the median weekly TypeScript check by 73%. Linear also rewrote type-aware lint rules as syntax-based analysis, reducing API lint time by 68% and full-repository lint time by 55%, then made the later Oxlint migration easier. Critical-path jobs were shortened by fetching only needed files, removing unnecessary checkouts, adding retrying checkout logic, and moving non-gating cache writes out of the merge path. Setup costs fell after shared dependencies were preinstalled, installs were limited to the API package, unnecessary caches were removed, schema snapshots replaced repeated migrations, and seven short checks were consolidated into two jobs; the batching change saved about 87,000 runner-minutes per month, or 11.8% of CI usage. With setup cheaper, Linear increased API testing from four to eight shards, making the critical job about 19% faster and cheaper in an initial benchmark. An opt-in Vitest configuration that shares module state produced the largest single saving, reducing the slowest shard from roughly 300-379 seconds to about 195 and cutting API-shard runner time from 32.8 to 22 minutes per run, although it required explicit eligibility and teardown because of correctness risks. Linear says these changes kept current test-suite time near five to six minutes instead of an estimated 11, while the codebase continues to gain about 2,000 tests per week.