What if the model learned its harness?

The hand-off from a coding model to its tools can consume calls before useful work begins. The Harness Hub article, published on November 6, 2026, puts that boundary inside the training problem: adapt the model to the harness that will receive and execute its actions.

The approach uses LoRA adapters for particular model-and-harness combinations, trained through reinforcement learning on verifiable tasks inside environments including pi, opencode and Hermes. The target is familiarity with the actual tools and conventions the assistant will encounter, without requiring a modified harness for the training setup.

The authors report held-out evaluations and fewer tool calls under examined conditions. Those findings should stay attached to the experiment. They are not a promise of efficiency across arbitrary repositories, and the discussion of reward exploitation highlights the difference between satisfying a measure and accomplishing the intended job.

I would assess the base model, adapter and harness as a working combination. When tools or instructions change, the fit needs checking again. Fewer calls matter only alongside preserved task quality. The operational choice is whether to keep that trained combination aligned with the workflow or accommodate the mismatch elsewhere in the agent.