Harness-Zero: Harness Distillation via Agent-as-Harness

Harness-Zero distills specialized agent-harness behavior into model weights, lifting macro-average task success from 23.3% to 44.3% without the harness at deployment.

Harness-Zero, listed on Hugging Face daily papers on 2026-09-20 and posted to arXiv on 2026-09-21, studies distilling the behavior induced by a domain- or instance-optimized agent harness into model weights so that a single fixed harness can be used at deployment. Because harnesses differ in action space and available information, a harnessing agent corrects or rewrites student responses into the target action space before execution, and those trajectories become fine-tuning demonstrations. Evaluated across knowledge work, tool use, and science tasks, macro-average success rises from 23.3% to 44.3% after the specialized harness is removed, exceeding the 41.7% achieved while the harness remains attached, and 82.3% of 28 induced behavior patterns are recovered. The Hugging Face report additionally notes that agent-as-harness beat code-as-harness for frontier LLMs using the same evolved harness. The two reports do not disagree; both give identical figures.

• Paper title: Harness-Zero: Harness Distillation via Agent-as-Harness.
• Listed on Hugging Face daily papers 2026-09-20 and posted to arXiv 2026-09-21.
• A harnessing agent rewrites student actions into the target action space, producing fine-tuning data.
• Macro-average success rose from 23.3% (specialized harness removed, no distillation) to 44.3% after distillation.
• The distilled 44.3% exceeds the 41.7% reached while the specialized harness remained attached.
• 82.3% of 28 harness-induced behavior patterns were recovered across three domains: knowledge work, tool use, and science.
• Agent-as-harness outperformed code-as-harness for frontier LLMs using the same evolved harness (Hugging Face report only).