Google open-sources Harness² framework for recursive agent self-improvement without retraining

A post says Google Research open-sourced Harness², a framework for improving the prompts, tools, memory and skills around an agent while leaving its model weights frozen. It describes task-level steps that propose, probe and compose harness changes, then carry lessons forward to future tasks. The post reports an 11.5-point gain on WorkBuddy-Bench.