Hugging Face has published a recipe for training small "harness adapters", showing that reinforcement learning inside an agent harness can make an open model both more accurate and cheaper to run on that harness. The post of November 20, 2026 lists Younes B, ben burtenshaw, Nathan Habib, Lysandre, Amine Dirhoussi, Adithya S K, shaun smith, Sergio Paniego and Joel Niklaus as authors.
The starting point is a complaint about general-purpose models: they are tuned for broad performance, so any particular combination of model and harness leaves accuracy and efficiency unused. The team trained one LoRA adapter per harness on top of Qwen3.8-27B, for the Pi, OpenCode and Hermes Agent coding harnesses, with the model frozen. For Pi on Terminal-Bench 2.1, Hugging Face reports that the base model scores 39 percent and uses about 2.9K tokens per task, while the adapted model reaches 55 percent pass@1 at about 1.5K tokens. It attributes most of the gain to the model reasoning less, calling fewer tools and getting more tasks right on the first try.
Training happens with the harness running exactly as shipped, rather than a re-implementation of the agent loop. A proxy between the harness and the vLLM server records the exact tokens produced, and the trainer learns from them using TRL's AsyncGRPOTrainer, OpenEnv and the Harbor task format, with datasets, training and rollouts all run as jobs on the Hugging Face Hub. The reward combines the verifier's pass or fail with a small 0.1 bonus for efficient tool use.
A large part of the post is about the tasks. The team searched the Hub for roughly 59,000 Harbor environments, kept sources built for training, and added tasks generated from merged pull requests with a pipeline called Tasksmith. Each task was checked with openenv validate: 858 of 1,127 Hub candidates passed. A difficulty check then kept tasks the base model solved only on some attempts. Of 442 scored tasks, the model solved 151 every time and 120 never, leaving 171 as useful training signal. The resulting dataset, harness-hub/agentic-coding-envs, holds 871 tasks from 35 repositories, and short training runs suggested that validation and difficulty filtering change what the model learns.
The authors also merged the three adapters into a single rank-96 super-adapter, applied at half strength. Across all three harnesses on Terminal-Bench 2.1 it scored 52.5 percent against 43.3 percent for the base model, with far fewer tool calls, though on some harness and benchmark pairs it fell behind the adapter trained for that harness alone. Applying the plain sum of the updates made agents too short and hurt Pi's score.
The adapters, GGUF files, RL tasks and a safety plugin are available in a Hugging Face collection. Every number comes from Hugging Face's own experiments with a single model family and one attempt per task, so independent reproduction is the next test of how far the recipe generalizes.