Linear reworked its CI pipeline and dropped median pull request wait from over 6 minutes to just over 5, on a test suite that quadrupled since January

Linear's engineering team published a post on 21 September 2026 saying its continuous integration pipeline had become the bottleneck after AI coding assistants started producing changes and tests faster than the CI could verify them. The team rewrote large parts of the pipeline and cut the median pull request wait from over 6 minutes to just over 5, on a test suite that has nearly quadrupled since January and that engineers and agents add 2,000 tests to every week. Without the rework, Linear projects the same suite would now take 11 minutes to run, double the current time.

The four changes that moved the numbers

The largest single win came from replacing tsc, the TypeScript compiler, with tsgo, the native Go rewrite Microsoft is building. Linear reports a 73 percent reduction in type-check time, with tsc workloads dropping 52 percent after the switch to a new runner provider that also gave a 34 percent speedup across other jobs on average. Full-repository lint went down 55 percent, and API lint 68 percent, on the same runners.

The critical path was next. Linear cut the change-detection job that decides which tests need to run from 26 seconds to 8 seconds at the median, and removed a cache-writing step from the merge path that saved 42 seconds. Per-shard setup came down from 110 to 140 seconds to 67 to 73 seconds, a 44 percent reduction, after preinstalling dependencies in the CI base image and skipping node_modules caching in favour of a faster reinstall, which ran in 7.5 seconds against the previous 28.

Test execution was reshuffled next. Sharding went from four to eight, and Linear split its largest test files to keep the load balanced across shards. A shared module-state cache, guarded by what the team calls "strict isolation rules", cut 17 percent of monthly rebuild time and pulled the average per-run cost from 300 to 379 seconds down to 195. Batching seven previously separate jobs saved 87,000 runner-minutes a month, which Linear puts at 11.8 percent of total CI usage.

Where AI coding actually pays off

AI coding delivers its speedup at the point where the review pipeline can confirm the code is correct. Every second the pipeline adds is a second the assistant spends idle waiting for a green build. Linear's fix is a stack of small mechanical wins, each measured, none glamorous. That order of operations, measure the slow step then delete it, is what a team can copy without knowing anything else about Linear.

None of these numbers translate directly to another team's setup. Linear runs a large TypeScript monorepo with its own test framework, and the tsgo switch alone is worth more to a codebase written mostly in TypeScript than to one in another language. The pattern that survives translation is the one at the top: measure the slowest step in the pipeline, cut it, then find the next one.

Source

Linear engineering, AI coding has made CI a bottleneck, so we reworked ours to keep up, 21 September 2026.