Watch the build on one of our agent PRs and most of what it does already happened ten minutes earlier. The agent that wrote the change ran tsc, ran biome for lint, ran the unit tests, and only then pushed. CI then did all three again from scratch, and took ten minutes over it.
With most of our commits written by agents[1], that is a lot of repetition. The agent has already run the checks. More to the point, it can prove it.
So now it does. The agent leaves a trailer under the commit message[2] recording what it ran before it pushed: tsc, biome, the unit tests.
The PR build reads the trailer and skips just those steps. Nothing else is skipped, and nothing else about the pipeline changed.
Builds went from about ten minutes to about a minute and a half, and this was a good third of that. The wall-clock saving was the obvious win, but the quieter one was contention: fewer Buildkite agents (the other kind of agent)[3] get spun up, fewer jobs sit waiting for one, and the ones that stay warm cache better. At 100+ PRs a day that is a lot of build minutes handed back.
The obvious question is trust, since a trailer is a claim, not a hash. Two answers. The trailer isn't the agent's word: a hook writes it, deterministically, when the checks actually run, so the agent can't skip one and claim it. And this is only the PR build: main re-runs everything, trailer or no trailer. For the record, zero bites so far.
This is a poor man's Bazel. Bazel does it properly, keying every action on a content hash and pulling the result from a remote cache[4]; this is the duct-tape version, keyed on a hook's record of what ran. It works because the agent did the work, the hook saw it, and the thing that matters gets checked again on main anyway.
Never Pick It Up Twice was about not touching a PR twice, and this is the same rule for the machines. If the agent has done the work, don't repeat it. Ask it what it ran.