harness-updating

This paper analyzes two capabilities in self-evolving LLM agents: harness-updating and harness-benefit. It finds that harness-updating is flat across base capability levels, while harness-benefit is non-monotonic, with mid-tier models benefiting most.