It's when LLM agents are inefficient that you need a whole rack of servers to get shit done.