1 paper
Yifan Zhang, Yutong Dai, Juntao Tan +9
An LLM agent's capability depends not only on model weights but on its harness: prompts, tools, skills, and control flow. Self-improvement loops already edit harnesses, yet single-…