Tag
This paper introduces Self-Harness, a new paradigm where LLM-based agents iteratively improve their own operating harness—prompts, tools, and control flow—without human engineers or stronger external agents, achieving significant performance gains across multiple models.