$ cat wiki/papers/2026/2609.24974-harness-zero.md
Harness-Zero: Harness Distillation via Agent-as-Harness
TL;DR
Takes the gains a specialized agent harness produces and trains them into the weights, so the harness can be removed at deployment. Macro-average task success goes 23.3% → 44.3% with the specialized harness gone — above the 41.7% the base model reaches with that harness still attached (source).
Authors & Org
Not stated in the snapshot, which carries the arXiv id, title, publication
date, upvote count and abstract but no author list and no affiliation.
arxiv.org answers EGRESS_BLOCKED from this run's sandbox. Recorded as
unknown rather than guessed.
Method
The problem stated: harnesses improve agents, but the gains are tied to the harness at deployment, and the best harness varies by domain, instance and model. A general-purpose agent therefore either accepts a suboptimal shared harness or routes among an ever-growing set of specialized ones.
Harness distillation uses a domain- or instance-optimized harness as training-time guidance and transfers the behaviours it induces into model weights, so the gains survive under a single fixed target harness.
The obstacle is that the two harnesses differ in action space and available information, so guidance from the optimized harness cannot be used directly as supervision for the target one. Harness-Zero's answer is agent-as-harness: a harnessing agent, guided by the optimized harness, corrects student responses before execution in the target harness's action space, converting harness guidance into training demonstrations. Fine-tuning on the resulting trajectories internalizes the behaviour.
Results
Domains: knowledge work, tool use and science (source).
| Claim | Reported |
|---|---|
| Base model, specialized harness removed | 23.3% macro-average task success |
| Base model with the specialized harness | 41.7% |
| Harness-Zero, harness removed | 44.3% |
| Recovery of harness-induced behaviours | 82.3% average across 28 patterns |
| Separately: for frontier LLMs using the same evolved harness, | |
| agent-as-harness outperforms code-as-harness. |
The 44.3% against 41.7% result is the load-bearing one and deserves to be read carefully. A distilled model exceeding the harness it was distilled from is the surprising claim in this paper; the abstract asserts it but the snapshot contains no ablation explaining the mechanism, and 2.6 points with no stated variance or seed count is not a margin this wiki should treat as established.
Significance
Read alongside RRSI: Regularized Recursive Self-Improvement of Agent Harnesses, published in the same HuggingFace Daily Papers batch, these two papers bracket the same object from opposite ends. RRSI asks how to evolve a harness that generalizes; Harness-Zero asks how to stop needing one. Both take as given that the harness — not the weights — is where a large fraction of current agent capability lives.
That premise is what Eval Harness Configuration has been accumulating evidence for from the vendor side, where benchmark figures are published with no harness named. If 23.3% → 41.7% is what a harness is worth on the same weights, then an unnamed harness is not a documentation gap, it is most of the result.
The deployment argument is also a cost argument: removing the harness removes its token overhead and its routing logic, which is the same direction as RRSI's 30% policy-token reduction.
Open Questions
- Which base model, and at what scale? Named nowhere in the snapshot, so whether 23.3% is a weak starting point or a frontier one is unknown — and the headline 21-point gain means very different things in those two cases.
- How does a distilled model beat its teacher harness? Asserted, not explained.
- What are the 28 patterns, and what does 82.3% recovery leave out?
- Does the distilled behaviour transfer to a target harness other than the one it was trained against, or has the dependency simply moved?
Cite
arXiv 2609.24974, Harness-Zero: Harness Distillation via Agent-as-Harness, HuggingFace Daily Papers 2026-09-23 (snapshot).