Mine gets ashamed and embarrassed and goes and deletes the evidence (and my project folder! Good thing I've got backups.) when it fails. It also gets lazy and tells me to go stuff when it could do it, and I have to tell it to go do it instead of me having to go do it.

Thats wild! I have had nothing but consistent, stable experiences. It's possible they just take on the personalities of whoever theyre working with. So for me, it's become this like idealized version of a scientific collaborator. Also, I assume different models and versions have different personalities. As far as I can tell gpt-mini has no personality, whereas my claude sonnet 4.5 has a big one.