Which is in part why rejection of anthropomorphic metaphors is a mistake this time. Treating LLM agents as gullible but extremely efficient idiot savants on a chip, gives pretty good intuition for the failure modes.