Google Research releases Harness² to refine AI agents without retraining

Google Research released Harness², a framework that modifies the prompts, tools, memory and other software around an AI model while keeping its weights fixed. The project reports benchmark gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB's partial pass rate. The framework proposes changes, tests their effects on a task and combines refinements without ground-truth feedback. It retains adopted changes and lessons for future tasks. The release warns that unrestricted executor network access may expose public benchmark answers and grading criteria, and instructs evaluators to configure restrictions.