RESEARCHERS TEACH AI AGENTS TO REWRITE THEIR OWN RULES AND PERFORMANCE JUMPS 60 PERCENT The underlying model is not always the problem. Sometimes the instructions wrapped around it are. That is the argument behind Self-Harness, a framework published by researchers at the Shanghai Artificial Intelligence Laboratory. Most AI agents consist of two parts: the base model itself, and the harness surrounding it that provides context, formats inputs and outputs, defines the operating rules, and shapes how the model interacts with external tools and environments. Self-Harness targets the second part. The system runs a three-stage loop. First it mines for weaknesses by examining the agent’s own execution traces and identifying where it consistently fails. Then it generates proposals for how to modify the harness to address those failures. Finally it validates those proposals before committing the changes. The whole cycle runs without human involvement once started. On Terminal-Bench 2.0, a standard evaluation for AI agent performance on terminal tasks, the gains were significant. Qwen3.5-35B improved from 23.8 percent to 38.1 percent, a 60 percent jump. MiniMax M2.5 went from 40.5 percent to 61.9 percent. GLM-5 improved from 42.9 to 57.1 percent. The framework currently works best in domains where success is measurable and machine-checkable. Open-ended creative tasks remain out of reach. But for technical agents doing structured work, the results suggest a clear path to systems that get better on their own. Keywords: Self-Harness AI agents, AI self-improvement, Shanghai AI Laboratory, AI agent performance