Google's Harness-Zero Boosts Agent Success Without a Harness

Agent harnesses do a lot of quiet work for AI models, and a new Google paper asks whether the model can just learn that work instead. In the paper's tests, the answer is often yes.

The number worth slowing down for: macro task success moves from 23.3% to 44.3% with the specialized harness removed, which is above the 41.7% the base model reaches with the harness attached.

What Happened

The method is called Harness-Zero, and the trick is timing. The optimized harness is used only during training, never at deployment.

That matters because the optimized harness and the deployment harness have different action spaces. A harnessing agent guided by the optimized harness corrects the student's responses in the deployment action space before those responses run. Those corrected runs then become the training demonstrations.

Across 28 harness-induced behaviors covering knowledge work, tool use, and science, 82.3% are recovered on average. For frontier models using the same evolved harness, the paper also finds that the agent-as-harness form beats the code-as-harness form.

Why It Matters

If this holds up, some of what a harness does today could be baked into the model rather than hand-written around it. That would mean less engineering per task and models that carry more of the workflow themselves.

The authors are upfront that robustness is still an open question. It is early work, but it points at something real: harness distillation is worth watching.