The unreasonable effectiveness of an LLM agent loop with tool use