Your CI is now the bottleneck

A 10-minute CI run used to be fine. If it took you two hours to write a change, ten more minutes waiting on checks was barely worth noticing.

Now an agent writes the same change in five minutes, and those ten minutes are two thirds of your cycle time. Your pipeline didn't get slower. Everything around it got faster.

This is Amdahl's law: speed up one part of a process and the parts you didn't speed up start to dominate. Programmers know it when it's about code, but we don't seem to apply it to our own work. For the last couple years most of the attention in AI coding has gone to models, prompts, and harnesses, which is to say to making the writing faster. Meanwhile the waiting lives in .github/workflows, which gets a fraction of that attention.

It's worse than the percentage suggests. When a person waits on CI, they do something else. An agent just waits, and if the checks fail it starts another ten-minute round to see whether its fix worked. Three tries to get green means half an hour of CI for five minutes of edits.

I think the deeper problem is that most pipelines were designed around how often humans push. If you push a few times a day, it's reasonable to run everything on every push, skimp on caching, and rerun flaky tests by hand. Each of those choices gets expensive when something pushes every few minutes. A flaky test that cost a person an occasional sigh now costs an agent a whole retry, and the agent can't tell the failure wasn't its fault.

So if you want your agents to ship faster, the best place to look may not be a new model. It may be your workflow file, and the question of which parts of it only made sense when a person was on the other end.