Harness-Zero Turns Agent Workflows Into Model Weights

A Peking University team says it has found a way to move part of an AI agent's scaffolding out of the serving stack and into the model itself. The approach, called Harness-Zero, distills behaviors learned through an agent harness into a smaller language model, potentially reducing the amount of orchestration code required at deployment.

What changed

As AI Weekly reports, Harness-Zero lifted a Qwen3.5-9B student's macro-average task success from 23.3% to 44.3% across SpreadsheetBench Verified, AppWorld, and USPTO Retrosynthesis. The distilled model also beat the 41.7% score recorded by the same base model when the specialized harness remained attached at inference.

The method uses a reviewer model to inspect the student's responses inside the target harness action space. Corrected trajectories then become supervised fine-tuning data. The paper reports 82.3% average recovery across 28 harness-induced behavior patterns in three domains.

Why builders should care

Agent systems often accumulate prompts, retries, planners, evaluators, tool wrappers, and memory logic around a foundation model. That scaffolding can improve reliability, but it also adds latency, token cost, operational dependencies, and more places for state to drift. If repeatable procedural behavior can be distilled into weights, part of that complexity becomes a training-time investment instead of a permanent serving-time tax.

The result is not a universal replacement for an agent harness. The gains were uneven: on USPTO Retrosynthesis, the distilled model scored 30.0%, while the meta-harness reached 38.0%. The researchers also report that trajectory collection ran 2.4 times slower than a vanilla mini-SWE-agent trial. Procedural behavior appears easier to internalize than deep domain knowledge.

The practical takeaway

For automation teams, the useful pattern is hybrid. Keep deterministic tools, permissions, approvals, and audit logs outside the model. Distill stable reasoning routines and domain-specific interaction patterns only after measuring them across real tasks. That can produce smaller, cheaper workers without pretending that model weights can replace execution controls.