AI coding agents made 25 changes—15 repaired or undid their own work

Twenty-five software changes in a day sounds like a productive team. The picture changes when fifteen of them repair or reverse work that team has just done.

That is the experience Julian Tellez described in a Bluesky thread on 6 November. He said he had given a group of AI agents a week on one of his products earlier in the year, with little design preparation. Of the 25 changes merged in one day, 15 fixed or undid the agents’ work from the same week. ¹

It is a personal account, without an independently checked project history. It is not a benchmark for AI coding systems, and the experiment itself is older than the post. Its interest lies in the distinction it opens between generating changes and deciding what the software should become.

Tellez argues that agent workflows can give extensive attention to building, testing and reviewing while leaving design underdeveloped. Where a requirement is unstated, the agent still has to make a choice. That choice can become somebody else’s correction later. ¹

The temptation is to read the count as a failure rate. It cannot tell us that. Fixes can be necessary, useful work, and human teams also revise their decisions. Nor do fifteen corrective changes necessarily consume more effort than the ten others.

But the account makes a useful dent in the appeal of activity figures. A busy change log can contain forward movement, maintenance and work created by earlier mistakes. Counting them together conceals the difference.

As AI makes producing software changes easier, the scarce contribution may increasingly be a clear decision about what deserves to change—and a reliable way to tell whether the result moved the product forward.