Google Unveils Harness2 to Boost Agent Performance Without Tweaking the Model

According to AI TIMES, Google Research on October 20 unveiled an open-source framework, Harness2, that boosts the performance of AI agents by letting them improve their own execution environment without changing model weights. In the released materials, the company said it confirmed improvements of up to 11.8 percentage points on professional-work benchmarks. The core idea is to repeatedly refine external factors such as prompts, tool usage methods, memory, and task procedures. The system is designed to compare differences in execution results even for tasks with no preset answers, identify directions for improvement, and reuse experience gained from one task as the starting point for the next. However, it is difficult to say this approach will directly translate into better real-world work quality, and the costs of repeated runs and the burden of verification also need to be considered. The release shows that the center of gravity in AI competition is shifting not only to larger models themselves, but also to the operating systems through which agents do their work.