It seems to me that Claude's error here (which is not unique to it) is self-sycophancy. The model is too eager to convince itself it did a good job.

I'd be curious to hear from experienced agent users if there is some AGENTS.md stuff to make the LLM more clear speaking? I wonder if that would impact the quality of work.