← In the News

Hugging Face trains one LoRA per coding harness and cuts tool calls by more than half

Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost · Lewis Tunstall, Nathan Habib and colleagues, Hugging Face · Hugging Face blog, November 20, 2026

Machine-readable Download Markdown

The team froze Qwen3.8-27B and trained a separate LoRA adapter for each of three agent harnesses, pi, Hermes Agent and OpenCode, with reinforcement learning run inside the unmodified harness. A proxy between the harness and the inference server recorded the exact tokens the model produced. The tasks came from SWE-smith, SWE-Gym and pull requests turned into tasks with a tool called Tasksmith. Each task was checked so that its reference solution earns full reward and wrong-answer controls do not, and those controls caught a Terminal-Bench 2.1 verifier that accepted known-wrong answers. On the 51 hardest SWE-bench Pro tasks, the pi adapter took pass@1 from 39.2% to 64.7% (15 tasks newly solved, 2 lost), Hermes went from 29.4% to 43.1%, and OpenCode from 35.3% to 51.0%. On tasks both models solved, the adapters made 56 to 72% fewer tool calls on SWE-bench Pro and 53 to 70% fewer on Terminal-Bench 2.1. Each result is one attempt per task, and the authors warn that scores cannot be compared across harnesses.

The authors flag their own open question. With thinking turned off, the adapted model writes short text before tool calls that "can sometimes look like reasoning traces," which they call reasoning leakage. On 300 business-automation tasks run in a different agent loop, the adapters stayed within noise of the base model, except perhaps the Hermes adapter, which fell 5.3 points. The adapters, the task dataset and all evaluation trajectories are public.

Why it matters: If you run an open-weight model in one harness, that harness can now be a training target. The published figures say a 27B model gets more correct answers in fewer steps after adaptation. Before relying on the efficiency figures, read the trajectories. The authors found the model writing reasoning-like text in a mode meant to skip reasoning, so part of the saving may be reasoning that moved rather than reasoning that disappeared.