Google open-sources Harness², boosting GLM-5 score by 11.5 points

ChainCatcher report: Google Research has open-sourced the AI toolchain self-improvement framework Harness², enabling AI agents to automatically modify system prompts, memory, skills, tool rules, and scripts or plugins without retraining models. The project is licensed under Apache 2.0 and supports OpenCode and Codex. Its core mechanism involves agents experimenting through trial and error: when encountering a new task, they propose several improvement strategies, which are then executed and evaluated by different agent versions, ultimately combining the most effective solutions. Even without a standard answer, the system identifies improvement directions based on actual execution outcomes. After task completion, the system retains information about which modifications were effective and why, for reference in future similar tasks. Google tested Harness² using GLM-5 and Gemini 3.5 Flash. On WorkBuddy-Bench, GLM-5 paired with OpenCode improved its average score from 64.2 to 75.7 after just one round of optimization, surpassing the unoptimized Gemini 3.5 Flash score of 72.4. In legal task evaluations, GLM-5’s pass rate on partial scoring criteria increased from 59.6% to 71.4%. In experiments, Claude Opus 4.8 was default responsible for proposing and implementing improvements, while GLM-5 and Gemini primarily executed tasks—not fully self-evolving. Domestic developer Joseph Zhou demonstrated integrating Harness² with Jev 1.14, reporting sustained improvements in visual task performance.