Some funny/sad glimpses of the Claude agent code.Querying the LLM to produce code is wasteful, even if it could one-shot it. But the takeaway from that thread is the agent is just constantly having the LLM retry, which is of course what an agent is supposed to do, which makes them extra wasteful by definition. But it really sinks in how "shot in the dark" the retry prompting is. There is definitely no science here, not even a little bit.
I think I saw somewhere that Dario Anthropic was proud that the agent was written by itself? Apologies if that is not true. But wow, it is waste producing waste.