Harness2 by Google Research – Recursive Agent Harnessing

Google Research has released Harness², a framework for optimizing AI agent systems through recursive refinement without changing model weights. The system uses two levels of recursion—task-level refinement that proposes and tests edits to improve performance on current tasks, and harnessing-level recursion that captures lessons learned to guide future improvements. Testing across three professional benchmarks showed significant performance gains, with improvements up to 11.5 points on WorkBuddy-Bench and 11.8 percentage points on LAB's pass rate for models like GLM-5 and Gemini 3.5 Flash.