I see a lot of "human type mistakes" in my LLM agent sessions. If you manage these human like mistakes with the same procedures that keep them under control in real humans, the LLMs become much more useful.