Google Research Releases Harness² as GLM-5 Average Score Gains 11.5 Points

Google Research released Harness², a framework that changes the software around AI models, as GLM-5’s average score on WorkBuddy-Bench climbed 11.5 points in one round without model retraining.

Running through OpenCode, GLM-5’s average score rose from 64.2 to 75.7. That was above the 72.4 score for Gemini 3.5 Flash using its unmodified setup on the same benchmark.

Harness² can revise system prompts, memory, skills, tool guidance, scripts and plugins. It tests proposed changes on a task, compares results and combines revisions that perform better. It retains effective changes and the reasons they worked for use on later similar tasks.

On legal tasks, GLM-5’s partial pass rate against scoring criteria increased from 59.6% to 71.4%. In the default setup, Claude Opus 4.8 proposed and implemented harness changes, while GLM-5 and Gemini 3.5 Flash mainly carried out the tasks.

The project is licensed under Apache 2.0 and supports OpenCode and Codex.