AI AGENTS CAN NOW REWRITE THEIR OWN RULES MID-TASK AND PERFORM UP TO 60 PERCENT BETTER AS A RESULT Researchers at the Shanghai Artificial Intelligence Laboratory published a new framework called Self-Harness that lets AI agents systematically revise their own operating rules based on what goes wrong during execution. The performance gains are significant enough to force a reconsideration of how enterprise AI workflows get optimized. The framework works by having an agent examine its own execution traces after completing tasks, identify patterns in where it fails, and propose targeted edits to its rule set. Those edits only get promoted if they demonstrably improve performance without introducing new regressions. The system trades manual prompt engineering from a human for empirical feedback from the agent’s own track record. Results across multiple models were consistent. MiniMax M2.5 improved from 40.5 to 61.9 percent on evaluation tasks, a 52.6 percent relative gain. Qwen3.5-35B-A3B went from 23.8 to 38.1 percent, a 60 percent gain. GLM-5 improved from 42.9 to 57.1 percent. What the improvements are not is just as important. They are not longer prompts or generic instruction padding. They are surgical edits tied to the specific failure patterns each model encounters in its actual operational environment. The enterprise implication is significant. AI agent performance today is largely a function of how well a human engineer wrote the original rules. Self-Harness suggests agents can take over a large portion of that optimization work themselves, without requiring any retraining of the underlying model. Keywords: Self-Harness AI framework, AI agent self-improvement, LLM performance optimization, enterprise AI agents