Researchers from Peking University, Google, and the Hong Kong University of Science and Technology have introduced Harness-Zero, a method that transfers the performance gains of a specialized AI agent harness into the model itself, so those gains no longer depend on external scaffolding. The work addresses a tension at the center of modern agent engineering: how much of an agent's capability should live in the model's weights, and how much in the scaffolding wrapped around it.

An agent harness is the external system that governs how a language model interacts with its environment — orchestrating tool use, managing context and state, and controlling the interaction loop. Harness engineering has proven to be a powerful lever for improving agent performance, but the improvements it delivers depend on the specific harness being present at deployment. Because the optimal harness varies across domains, individual tasks, and base models, a general-purpose agent faces a difficult trade-off: adopt a single shared harness and forfeit specialized advantages, or maintain a growing collection of specialized harnesses with mounting costs in routing, context, and orchestration.

Harness-Zero proposes a third path: harness distillation. An optimized harness serves only as training-time guidance, and the behaviors it induces are transferred into the model's weights, so the gains persist even after the specialized harness is removed and a minimal fixed harness is used at deployment.

Agent-as-Harness: Translating Guidance Across Different Action Spaces

The central difficulty is that the optimized harness and the target harness differ in both action space and available information, meaning guidance from the optimized harness cannot be applied directly as training supervision. Harness-Zero addresses this with an agent-as-harness approach: a separate harnessing agent, guided by the optimized harness, reviews each response the student agent proposes and, where necessary, corrects it so that the correction is expressed in the target harness's native action space before execution.

The corrected runs then serve as training demonstrations. Fine-tuning on the resulting trajectories internalizes the harness-induced behavior directly in the model's weights, and at deployment the specialized harness, the reference harness, and the harnessing agent are all discarded.

The researchers evaluated the method across three domains — spreadsheet-based knowledge work (SpreadsheetBench Verified), multi-application tool use (AppWorld), and scientific reasoning (USPTO Retrosynthesis). In training-free evaluations on frontier models — that is, with no fine-tuning involved — the agent-as-harness approach outperformed the conventional code-as-harness method, averaging 81.1% versus 78.1% across the benchmark and model settings tested.

For distillation, the macro-average task success of a 9-billion-parameter base model, comparatively compact by current standards, rose from 23.3% to 44.3% once the specialized harness was removed — exceeding the 41.7% the same base model achieved with the harness still attached. A behavioral analysis further found that the distilled model recovered an average of 82.3% of 28 harness-induced behavior patterns that were absent from the base model, an indication that the distilled model absorbed how the harness worked, not merely how well it performed.

The authors note several limitations. The approach requires a sufficiently capable model to serve as the harnessing agent, since weaker models produce net-harmful interventions, and the review process raises the cost of trajectory collection. Deeper domain knowledge encoded in a harness may also be harder to internalize through fine-tuning alone.

Even so, the results suggest that harness development could become a scalable source of training signal — with better models building better harnesses, and each harness returning its gains to the model itself rather than leaving them locked in the scaffolding around it. How far that loop can stretch, given the constraints the authors flag, is the open question that will shape whether harness gains keep landing in the scaffolding — or start landing in the weights themselves.