A senior developer stopped using AI at work for a month and got smaller, better-tested pull requests back

Anyone measuring engineering output by pull requests per day is measuring something an AI agent can fake overnight. A developer who blogs as Bustikiller, and who describes himself as a test-driven-development practitioner of ten-plus years, published a note on September 25 about stopping AI use at work for a month, and it reached 167 points on Hacker News within eleven hours. The post is called One month without AI, and it is a first-person account of what a senior engineer thinks he lost while looking productive.

What the setup looked like

He describes running several coding agents at once, in separate git worktrees, connected to Jira through the atlassian command-line client so the agents could pick up tickets by ID. He would paste a Jira description in and let the agent build the pull request. Small tasks came back in about thirty minutes and he used the wait to switch to another task. Larger tasks were decomposed by the same agent and returned as convincing pull requests with seven-paragraph descriptions, which he stopped reading. He turned off the setting that co-signed his commits because he did not want the AI's name on them.

Where it broke

A coworker caught it. In a review of one of Bustikiller's pull requests, the coworker pointed out that a test the AI had written did not actually cover the scenario the change was for. He reread it and agreed. He had been a TDD engineer for a decade, and the test had been convincing enough that he had let it through. He also recalls one agent stalling on a task for thirty minutes and running up about $30 in token cost before he told it off in the chat and got the answer in twenty seconds. He writes that his estimate of how much of the code in his own pull requests he actually understood had dropped well below twenty percent, and that friends in the same company had said the same.

What changed when he stopped

The pull requests shrank to about five files. The descriptions went back to two lines. He could answer questions in code review, because he had written the change and could justify each line. Tests actually tested the scenario the change concerned. He does not claim he ships more than he did, and he acknowledges that pull-request count and line count are metrics AI can beat by being verbose. His claim is narrower: that he could no longer notice the difference between good and bad code produced by the agent, and that the loss of that judgment was the thing worth stopping to prevent.

The post is worth reading in full for anyone whose team runs multiple agents on Jira tickets and calls the result throughput. It is one person's account, not a study, and one anecdote is not a rule. The pattern it describes is now common enough that the piece drew wide agreement on a public forum in the middle of a workday, and the failure it names is invisible from the outside: an experienced engineer's pull requests looked fine, cleared review most of the time, and were slowly leaving him unable to defend them. A team relying on volume of AI-drafted work as a productivity signal is looking at the wrong number.

Source

Primary source: One month without AI on Bustikiller's Blog, September 25, 2026. Discussion: Hacker News item 49855018.