metaevo-ai / harness-zero

# Harness-Zero: Harness Distillation via Agent-as-Harness
**Distill an optimized agent harness into model weights — deploy with a minimal harness, keep the gains.**
Agent harnesses — the external systems that mediate model–environment interaction — can substantially improve agent performance, but their gains remain tied to the harness at deployment. **Harness-Zero** transfers the behaviors an optimized harness induces into the model itself: a *harnessing agent* reviews the student's every response during training-time rollouts and rewrites it, in the student's own action space, whenever the optimized harness would have done better. Fine-tuning on the reviewed trajectories internalizes the behavior — at deployment the harness is gone and the gains remain. The full paper is available on alphaXiv.
## Highlights
- 📈 **+21.0 points macro-average** on the base model (23.3% → 44.3%) across three domains — a 90.1% relative improvement.
- 🏆 The distilled model **under the minimal harness alone (44.3%) beats the base model with the optimized harness still attached (41.7%)**.
- 🔁 **Agent-as-harness > code-as-harness**: the same evolved harness applied through a harnessing agent averages 81.1% vs. 78.1% for direct mounting on frontier models.
- 🧠 **82.3% recovery** of 28 harness-exclusive behavior patterns (memory, skill, tool, middleware) — direct evidence the behavior moved into the weights.
## Results
Distillation into Qwen3.5-9B:
| Setting | SpreadsheetBench | AppWorld | USPTO | **Avg.** |
|---|---|---|---|---|
| mini-SWE-agent (target harness *h*) | 31.0 | 26.8 | 12.0 | 23.3 |
| meta-harness (evolved harness *h\**) | 39.0 | 48.2 | **38.0** | 41.7 |
| DeepAgents | 35.0 | 19.6 | 7.0 | 20.5 |
| Claude Code | 31.0 | 10.7 | 6.0 | 15.9 |
| **Harness-Zero (distilled, *h* only)** | **44.0** | **58.9** | 30.0 | **44.3** |
## How it works
1. **Evolve & Adapt.** A student-side harness *h\** (tools, middleware, skills, memory —
 the DeepAgents abstraction) is evolved on training tasks, then adapted into a private reference harness *K* (action recipes, review middleware, review guidance, failure patterns) for the harnessing agent. See `harness_bank/` for the evolved harnesses and `.agents/skills/` for the evolution skill.
2. **Agent-as-harness trajectory collection.** The student runs under a fixed, minimal
 mini-SWE-agent-style target harness *h* (one bash tool, fixed system prompt). Before any response is executed, the harnessing agent reviews it using *K* and either **PASS**es it unchanged or **REPLACE**s it with a complete response valid under *h*. Only the accepted response enters the student-visible trajectory; the review stays private.
3. **Train & deploy.** SFT on the reviewed trajectories (masking any reviewer-perspective
 reasoning). The distilled student deploys under *h* alone — no *h\**, no *K*, no harnessing agent.
## Installation
 [code] 
Requirements: Python 3.12–3.13, Docker (rollouts run in local Docker sandboxes via Harbor).
**Environment variable