Google open-sources Harness²: Allows AI to autonomously improve toolchains, lifting GLM-5's benchmark score by 11.5 points.

Insight: Beating AI Flash News, Google Research has open-sourced Harness², enabling AI agents to automatically modify system prompts, memory, skills, tool rules, and even scripts and plugins without retraining the underlying model. The project is licensed under Apache 2.0 and supports OpenCode and Codex. Its core mechanism lets agents learn through trial and error: when facing a new task, the system first proposes multiple improvement plans, has different agent versions execute them separately, compares their performance, and finally combines the best elements to form a more suitable solution. Even without a standard answer, it can identify improvement directions based on actual execution processes. After a task is completed, the system retains records of which modifications worked and why, so it can reference this data for similar tasks in the future instead of starting from scratch each time. Google tested Harness² using GLM-5 and Gemini 3.5 Flash. On WorkBuddy-Bench, GLM-5 paired with OpenCode saw its average score jump from 64.2 to 75.7 after just one round of optimization, outperforming unoptimized Gemini 3.5 Flash’s 72.4 points. In legal task tests, GLM-5’s pass rate for some scoring criteria also rose from 59.6% to 71.4%. However, the experiment defaulted to Claude Opus 4.8 for proposing and implementing improvements, with GLM-5 and Gemini primarily responsible for executing tasks—meaning the evolution was not achieved entirely by these models on their own. Chinese developer Joseph Zhou also demonstrated an experiment integrating Harness² with Jev 1.14, noting consistent improvements in performance on visual tasks.