{
  "schema_version": "1.1",
  "id": "s13:https://arxiv.org/abs/2609.09134v1",
  "slug": "2609-09134v1-0727nfl",
  "url": "https://feed7.dev/p/2609-09134v1-0727nfl",
  "title": "Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails",
  "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.",
  "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_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.\n\nTreat 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.\n\nThe 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.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2609.09134v1",
    "published_at": "2026-09-08T17:53:49.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "agent",
  "domains": [
    "coding"
  ],
  "topics": [
    "harness-engineering",
    "agent-reliability",
    "enterprise"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "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."
  ],
  "connected_context": {
    "meaning": "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.",
    "corpus_size": 713,
    "generated_at": "2026-09-09T10:12:53.401Z",
    "connections": [
      {
        "title": "OpenForgeRL: Train Harness-native Agents in Any Environment",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.21557v1",
        "feed7_url": "https://feed7.dev/p/2607-21557v1-0blvz16",
        "reason": "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."
      },
      {
        "title": "The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.24720v1",
        "feed7_url": "https://feed7.dev/p/2607-24720v1-0gihy13",
        "reason": "The degradation from full expert imitation reinforces the earlier finding that teacher-pattern compatibility matters for long-horizon planning."
      },
      {
        "title": "SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2609.02786v1",
        "feed7_url": "https://feed7.dev/p/2609-02786v1-0eo17gt",
        "reason": "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."
      },
      {
        "title": "Procedural Graphs: Self-Evolving Execution Structures for LLM Agents",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2609.09153v1",
        "feed7_url": "https://feed7.dev/p/2609-09153v1-0ee5zbi",
        "reason": "Procedural Graphs evolve execution guidance from trajectories; this result implies that such harness evolution should be evaluated jointly with any subsequent model-weight adaptation."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-09-08T17:53:49.000Z",
  "modified_at": "2026-09-08T17:53:49.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2609-09134v1-0727nfl",
    "json": "https://feed7.dev/p/2609-09134v1-0727nfl.json",
    "markdown": "https://feed7.dev/p/2609-09134v1-0727nfl.md"
  }
}