# Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

Source: [arXiv](https://arxiv.org/abs/2609.09134v1)  
Feed7 permalink: https://feed7.dev/p/2609-09134v1-0727nfl  
Published: 2026-09-08T17:53:49.000Z  
Trust: Needs Review (needs_review)

## Why Included

Full expert-trajectory imitation can break the fit between a weaker model and its evolved harness. Correcting only the failing turn preserved the model’s planning style across seven enterprise tasks.

## Source Summary

Across **seven enterprise tasks**, full expert-trajectory training under an evolved harness reduced weaker-model performance by **4–30 points** for Qwen3-Coder and Gemma 4. The same imitation procedure helped when the harness had not been evolved.

## Practical Implication

Treat model and harness changes as a coupled system. If a harness was optimized around a model’s native planning style, adapt from that model’s own rollouts and use expert correction only at the localized failing turn.

## Agent-Ready Context

Across **seven enterprise tasks**, full expert-trajectory training under an evolved harness reduced weaker-model performance by **4–30 points** for Qwen3-Coder and Gemma 4. The same imitation procedure helped when the harness had not been evolved.

Treat model and harness changes as a coupled system. If a harness was optimized around a model’s native planning style, adapt from that model’s own rollouts and use expert correction only at the localized failing turn.

The proposed on-policy pipeline preserves planning style and combines harness and weight adaptation in these experiments. Its evidence is limited to the studied enterprise tasks and models, so broader transfer remains open.

## Connected Context

Feed7 judgment across 713 accumulated Signals:

This narrows harness-native training: matching the deployment environment is insufficient if adaptation overwrites the model’s native planning style. The reported reversal under an evolved harness makes model and harness versions a coupled evaluation unit and favors corrections to the model’s own failing turns over wholesale imitation of expert trajectories.

- [OpenForgeRL: Train Harness-native Agents in Any Environment](https://feed7.dev/p/2607-21557v1-0blvz16) — OpenForgeRL establishes training inside the deployment harness; this result adds that such training should remain on-policy and preserve the deployed model’s planning style.
- [The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation](https://feed7.dev/p/2607-24720v1-0gihy13) — The degradation from full expert imitation reinforces the earlier finding that teacher-pattern compatibility matters for long-horizon planning.
- [SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment](https://feed7.dev/p/2609-02786v1-0eo17gt) — Both couple harness evolution with policy adaptation from agent experience, but this work specifically shows why localized on-policy correction can be safer than copying full expert trajectories.
- [Procedural Graphs: Self-Evolving Execution Structures for LLM Agents](https://feed7.dev/p/2609-09153v1-0ee5zbi) — Procedural Graphs evolve execution guidance from trajectories; this result implies that such harness evolution should be evaluated jointly with any subsequent model-weight adaptation.

## Context Map

- Layer: agent
- Domains: coding
- Topics: harness-engineering, agent-reliability, enterprise

## Uncertainty

- The proposed on-policy pipeline preserves planning style and combines harness and weight adaptation in these experiments. Its evidence is limited to the studied enterprise tasks and models, so broader transfer remains open.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.