Harness-Zero trains harness behaviour into weights, then removes the harness

single source· 1 articles · confidence: medium · first seen 2026-09-20 20:00 UTC

What this means for you

If you maintain an agent harness, this is a direction to watch rather than adopt: no code, weights or model names are given, so the 23.3% to 44.3% gap cannot be checked. The claim worth taking seriously is that harness gains may be trainable into weights rather than engineered per deployment.

Harness-Zero is a training method that transfers the benefit of an agent harness — the scaffolding that lets a model call tools and act on its environment — into the model's weights, so the specialised harness can be dropped at deployment. An optimised harness guides an agent that rewrites the student's answers in the target harness's action space; those corrected trajectories become training data. Across knowledge work, tool use and science tasks the authors report macro-average success rising from 23.3% to 44.3%, above the 41.7% reached with the specialised harness attached, and 82.3% recovery across 28 behaviours. It is a preprint: the models, evaluation dates and benchmarks are not named.

Key facts

  • ·Harness-Zero reports macro-average task success of 44.3% with the specialised harness removed at deployment, against a 23.3% base model. source
  • ·The same base model reaches 41.7% macro-average task success with the specialised harness still attached, below the 44.3% reported after distillation. source
  • ·The method recovers 82.3% of harness-induced behaviours on average across 28 patterns spanning knowledge work, tool use and science. source
  • ·For frontier LLMs using the same evolved harness, agent-as-harness outperformed code-as-harness in the reported experiments. source
  • ·The work is posted as arXiv preprint 2609.24974 dated 20 September 2026; the abstract does not name the models, benchmarks or evaluation dates. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire