Hugging Face shows how small adapters tune open models to your coding agent
Hugging Face trained RL adapters that make Qwen3.8-27B more accurate and use fewer tokens inside Pi, OpenCode and Hermes Agent, and published the full recipe.
Hugging Face published a recipe on November 20, 2026 for what it calls harness adapters: small add-on weights, trained with reinforcement learning inside a specific coding agent, that make an open model work better and more cheaply in that agent. The team trained one LoRA adapter for Qwen3.8-27B for each of three harnesses: Pi, OpenCode and Hermes Agent.
What's new
- Training happens in the real tool: instead of re-creating an agent loop inside a training framework, Hugging Face says the harness runs exactly as it ships, and a proxy between the harness and the inference server records the tokens used for training.
- Headline result: the blog says Qwen3.8-27B scored 39% on Terminal-Bench 2.1 under the Pi harness while using about 2.9K tokens per task. With its weights frozen and an RL-trained adapter added, it reached 55% pass@1 at roughly 1.5K tokens per task.
- Why it improves: according to the authors, most of the gain comes from the model reasoning less, making fewer tool calls and getting tasks right on the first attempt.
- One adapter for all three: the three LoRAs were also merged into a single merged super adapter that the team says runs unchanged under Pi, Hermes Agent and OpenCode.
- Everything is public: adapters, GGUF builds, the RL tasks and a safety plugin are shared in a Hugging Face collection, alongside the full training pipeline.
- Open question: the authors note that in non-thinking mode the adapted model tends to write short reasoning-like text before tool calls. They say the results still hold under the same token budget but plan follow-up work on thinking models.
Why it matters
Analysis: most people pick a coding agent and a model separately and accept whatever fit they get. This work suggests a cheap middle path: keep the open model you already run and add a small adapter tuned to the agent you actually use, which can cut token costs as well as raise success rates. The results come from Hugging Face's own tests on two benchmarks, so teams should check them against their own tasks before relying on them.
Source: Hugging Face