Harness-Zero: Harness Distillation via Agent-as-Harness
Harness-Zero distills specialized agent harnesses into weights, lifting task success from 23.3% to 44.3%.
Harness-Zero distills behavior induced by a domain- or instance-optimized agent harness into model weights so a single fixed harness can be used at deployment. A harnessing agent rewrites student actions into the target action space, creating fine-tuning demonstrations when the harnesses differ. Across knowledge work, tool use, and science, agent-as-harness beat code-as-harness for frontier LLMs using the same evolved harness. Removing the specialized harness raised macro-average success from 23.3% to 44.3%, above the 41.7% with it attached, and recovered 82.3% of 28 induced patterns.
- Distills an optimized harness into weights for one fixed target harness.
- Agent-as-harness outperformed code-as-harness for frontier LLMs.
- Macro-average success rose from 23.3% to 44.3% after harness removal.
- Recovered 82.3% of 28 harness-induced patterns across three domains.
Full article248 words · extracted from huggingface.co · click to collapse
Agent harnesses, the external systems that mediate model-environment interaction, can substantially improve agent performance, but their gains remain tied to the harness at deployment. Because the best harness varies across domains, instances, and models, a general-purpose agent must either settle for a suboptimal shared harness or route among an ever-growing set of specialized ones. We therefore study agent harness distillation: using a domain- or instance-optimized harness as training-time guidance and transferring the behaviors it induces into model weights, so that its gains survive under a single fixed target harness. The challenge is that the two harnesses differ in action space and available information, so guidance from the optimized harness cannot serve directly as supervision for the target one. We introduce Harness-Zero, which enables harness distillation through agent-as-harness. Guided by the optimized harness, a harnessing agent corrects student responses before execution in the target harness's action space, turning harness guidance into training demonstrations. Fine-tuning on the resulting trajectories internalizes harness-induced behavior into the model, so the specialized harness can be removed at deployment. Our experiments spanning knowledge work, tool use, and science domains show that: (1) For frontier LLMs using the same evolved harness, agent-as-harness outperforms code-as-harness. (2) With the specialized harness removed at deployment, Harness-Zero improves the base model's macro-average task success from 23.3% to 44.3%, even exceeding the 41.7% it reaches with that harness still attached. (3) Harness-Zero recovers harness-induced behaviors absent from the base model, with 82.3% average recovery across 28 patterns in the three domains.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.24974