Paper Overview
Field: NLP Authors: Haoran Ye, Yuxing Lu, Haonan Dong, Zhaochen Su, Guojie Song Published: 2026-09-21 arXiv: 2609.24974
Summary
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.
The authors 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 harness.
Harness-Zero addresses this via Agent-as-Harness: guided by the optimized harness, a harnessing agent corrects the student model's responses within the target harness's action space before execution, turning harness guidance into training demonstrations. Fine-tuning on the resulting trajectories internalizes harness-induced behaviors into model capabilities, removing the specialized harness at deployment.
Key Findings
Experiments span knowledge work, tool use, and science domains:
1. Agent-as-Harness beats Code-as-Harness for frontier LLMs using the same evolving harness. 2. Deployment without the specialized harness: Harness-Zero lifts the base model's macro-average task success rate from 23.3% to 44.3%, even surpassing the 41.7% achieved when the harness was still present. 3. Behavior recovery: Harness-Zero restores harness-induced behaviors missing from the base model, recovering on average 82.3% across 28 patterns in three domains.
Links
- Paper: https://arxiv.org/abs/2609.24974