Self-Harness: Harnesses That Improve Themselves (arxiv.org)

A groundbreaking study introduces "Self-Harness," a novel approach allowing large language model (LLM)-based agents to autonomously improve their own operational harnesses, without human intervention. This paradigm addresses the challenge of model-specific harness design, which has traditionally relied on expert engineers, a method that struggles to keep pace with the rapidly evolving landscape of diverse LLMs. Self-Harness operates through an iterative three-stage process: Weakness Mining identifies model-specific failures, Harness Proposal generates tailored modifications, and Proposal Validation rigorously tests these changes to ensure performance enhancement. In trials with Terminal-Bench-2.0 utilizing three distinct LLMs, Self-Harness significantly boosted performance metrics, with pass rates improving from 40.5% to 61.9% for MiniMax M2.5, 23.8% to 38.1% for Qwen3.5-35B-A3B, and 42.9% to 57.1% for GLM-5. The paradigm demonstrates that rather than applying generic adjustments, Self-Harness transforms specific weaknesses into actionable modifications, paving the way for LLM agents that not only operate effectively but can also iterate on their own structures. This advancement heralds a shift toward more adaptive and intelligent AI systems capable of self-optimization, significantly impacting the AI/ML community's approach to harness development.