Google Research releases Harness2 to rewrite agent control software
One change to the harness increased the GLM-5 score by 11.5 points, with no change to the model weights.
Claimed, not confirmed
Google Research released Harness2, a free framework with the Apache-2.0 licence. It rewrites the harness of an agent while the agent runs. The harness is the control software around a model. One change to it increased the GLM-5 score by 11.5 points on WorkBuddy-Bench. The company gave no score from before the change and no other test. No other group has checked the result.
Sources
Posted
