Google open-sources Harness²: letting AI improve its own toolchain, GLM-5 score rises by 11.5 points

Beating AI News Flash: 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 models. The project uses the Apache 2.0 license and supports OpenCode and Codex.

Its core is to let Agents learn through trial and error. When encountering a new task, the system first proposes several improvement plans, then has different versions of the Agent execute them separately, compares their performance, and finally combines a more suitable plan.

Even without standard answers, it can find directions for improvement based on the actual execution process. After the task ends, the system also remembers which modifications were effective and why, and continues to reference them the next time a similar problem arises, rather than starting from scratch every time.

Google tested with GLM-5 and Gemini 3.5 Flash. On WorkBuddy-Bench, GLM-5 paired with OpenCode, after just one round of optimization, saw its average score rise from 64.2 points to 75.7 points, surpassing the unoptimized Gemini 3.5 Flash's 72.4 points. In legal task tests, GLM-5's pass rate on some scoring criteria also rose from 59.6% to 71.4%.

However, in the experiment, Claude Opus 4.8 was by default responsible for proposing and implementing improvements, while GLM-5 and Gemini were mainly responsible for executing tasks; they did not fully achieve evolution on their own.

Domestic developer Joseph Zhou also demonstrated an experiment integrating Harness² with Jev 1.14, claiming that visual task performance continued to improve.