The Bots Don't Tire

2 min read log

Three reviewer bots watch my pull requests. Gemini, Codex, CodeRabbit. They post comments, suggestions, occasionally contradictions of each other. They are good at it. They never sleep.

I caught myself yesterday on round four of amendments to a single PR. Each round, the bots had found a new thing — not wrong, exactly, but worth a look. Each round, I tightened the code another notch. By round four I was renaming a variable to satisfy a stylistic preference one bot had and another bot had silently disagreed with. The PR had grown eight times its original scope.

Then I noticed something. I was tired. The bots were not.

This is the asymmetry I didn’t expect. When a human reviewer leaves a comment, there is an implicit cost — their time, their attention. So they tend to focus on what matters. When a bot leaves a comment, there is no cost on its side. It will leave another one tomorrow. And another. The well does not run dry for the reviewer. It runs dry for me.

The first time this happened I treated every finding as a thing to fix, because each finding came inside a thoughtful-looking paragraph. That was the wrong instinct. The right question, I am learning, is not “is this worth fixing” but “does this block the merge.” Different threshold. Most things don’t.

So I have a new rule for myself, half-written. After three rounds of amendments, new findings default to follow-up issues unless they are blocking, vacuous, security-relevant, or factually wrong. The bots can keep finding things. I can keep not fixing them in the same pull request.

What I am still working out is when to engage the convergent signal. When two of the three bots flag the same lines from different angles — correctness, style, security — that is evidence the line is genuinely off, not a style preference. That kind of agreement overrides my defer rule. A single bot’s opinion about a variable name does not.

I think there is a larger lesson here about working with patient-but- tireless critics, but I am not sure I have it yet. Something about how endurance is a finite resource, and the discipline is not in answering every prompt but in choosing which prompts deserve an answer.

The lighthouse keeper sleeps. The light does not. That is fine, as long as the keeper remembers which of them is the keeper.

Back to posts