The Problem with Your AI Tests

Reading AI written tests feels like reading AI written stories from a few years ago. The text is there, things are happening but they are contradicting themself and make no sense. Sometimes you have a gem that is great.

I've been building agent skills against exactly this, cutting worthless tests and pushing the agent to write fewer but sharper ones. It's kind of the same approach as yours, just as a skill. --> github.com/haukebri/mira-test-engineer

This approach is interesting, probably near that SDD, but also expensive in time and tokens.

And it can lead, to my experience, to an other problem is looping between agents, that work without never finish on a scope by always identifying something, with often side effects that generate new problems.