I 100% agree with you, which is why I try to be polite to the LLMs. Also, on a practical note, I've never seen an agent session go right-side-up after I crash out; the model seems to perform worse than if I just gently guide it.