Researchers introduced Harness-Zero to distill agent harness guidance into model weights. It uses an agent to correct student responses before execution. Fine-tuning internalizes harness-induced behavior into the model. Experiments covered knowledge work, tool use and science domains. Agent-as-harness outperformed code-as-harness for frontier models. Deployment success rose from 23.3% to 44.3% without the specialized harness. This exceeded the 41.7% success achieved with the harness still attached. Harness-Zero recovered 82.3% of harness-induced behaviors on average. The paper was submitted to arXiv on 21 September 2026.
What changed
Agent gains were tied to the harness used at deployment.
What it unlocks
Removing specialized harnesses at deployment while keeping their performance gains.
- 23.3% to 44.3% macro-average task success
- 41.7% with harness attached
- 82.3% average behavior recovery
Sources