Researchers explore harness evolution for long-horizon LLM agents, showing that self-evolving harnesses can improve model performance and that trained weights can internalize gains from harness evolution. This work demonstrates a diagnose-then-intervene approach to choosing between harness and weights.