Harness-Zero reportedly improves task success without a specialized agent harness
A user sharing a paper from Google and colleagues describes a way to train agents with an optimized harness—the setup guiding their behavior—then remove that specialized setup for deployment.
TLDR
According to the user's account, Harness-Zero uses an optimized harness only during training. A harness-guided agent corrects the student's responses to fit the actions available at deployment before they run; those corrected runs become training demonstrations. The post reports macro task success rising from 23.3% to 44.3% without the specialized harness, exceeding the base model's 41.7% with the harness attached. It also reports that 82.3% of 28 harness-induced behaviors across knowledge work, tool use and science are recovered on average. The user cautions that the approach's robustness remains to be seen.
