Harness-Zero: Google Researchers Demonstrate How To Distill An AI Agent Harness Into Model Weights

by
Alisa Davidson

Revealed: September 25, 2026 at 8:07 am Up to date: September 25, 2026 at 8:07 am

by Anastasiia O

Edited and fact-checked:
September 25, 2026 at 8:07 am

To enhance your local-language expertise, generally we make use of an auto-translation plugin. Please observe auto-translation will not be correct, so learn original article for exact info.

A workforce of researchers from Peking University, Google, and the Hong Kong University of Science and Technology has launched a technique referred to as Harness-Zero, which permits the efficiency good points of a specialised AI agent harness to be transferred right into a mannequin itself reasonably than remaining tied to exterior scaffolding.

An agent harness is the exterior system that manages how a language mannequin interacts with its atmosphere — orchestrating instrument use, managing context and state, and controlling the interplay loop. Harness engineering has confirmed to be a robust lever for bettering agent efficiency, however the good points it produces depend upon that particular harness being current at deployment. As a result of the optimum harness varies throughout domains, particular person duties, and base fashions, a general-purpose agent faces a tough trade-off: both settle for a single shared harness that forfeits specialised benefits, or preserve a rising assortment of specialised ones with the related prices in routing, context, and orchestration.

Harness-Zero proposes a 3rd path: harness distillation. An optimized harness is used solely as training-time steerage, and the behaviors it induces are transferred into the mannequin’s weights, in order that the good points persist even when the specialised harness is eliminated and a minimal mounted harness is used at deployment.

Agent-as-Harness: Translating Steerage Throughout Totally different Motion Areas

The central issue is that the optimized harness and the goal harness differ in motion area and out there info, which means steerage from the optimized harness can’t be used instantly as coaching supervision. Harness-Zero solves this via an agent-as-harness strategy: a separate harnessing agent, guided by the optimized harness, evaluations every response the coed agent proposes and, when essential, corrects it so the correction is expressed within the goal harness’s native motion area earlier than execution. 

These corrected runs develop into the coaching demonstrations. Superb-tuning on the ensuing trajectories internalizes the harness-induced habits within the mannequin, and the specialised harness, the reference harness, and the harnessing agent are all discarded at deployment.

The researchers evaluated the strategy throughout three domains — spreadsheet-based data work (SpreadsheetBench Verified), multi-application instrument use (AppWorld), and scientific reasoning (USPTO Retrosynthesis). In training-free evaluations on frontier fashions, agent-as-harness outperformed the traditional code-as-harness strategy, averaging 81.1% versus 78.1% throughout the benchmark and mannequin settings examined.

For distillation, a 9-billion-parameter base mannequin’s macro-average process success rose from 23.3% to 44.3% with the specialised harness eliminated — exceeding the 41.7% the bottom mannequin achieved with the harness nonetheless connected. A behavioral evaluation discovered the distilled mannequin recovered a median of 82.3% of 28 harness-induced habits patterns that had been absent from the bottom mannequin.

The authors observe limitations: the strategy requires a sufficiently succesful mannequin to function the harnessing agent, since weaker fashions produce net-harmful interventions, and the evaluation course of will increase the price of trajectory assortment. Deeper area data encoded in a harness may additionally be tougher to internalize via fine-tuning alone. Nonetheless, the outcomes recommend harness improvement might develop into a scalable supply of coaching sign — with higher fashions constructing higher harnesses, and every harness returning its good points to the mannequin.

Alisa, a devoted journalist on the MPost, makes a speciality of crypto, AI, investments, and the expansive realm of Web3. With a eager eye for rising developments and applied sciences, she delivers complete protection to tell and have interaction readers within the ever-evolving panorama of digital finance.

• 
  
  
  

More articles