Hugging Face has published a method for RL training of LoRA adapters for agentic harnesses
Hugging Face has published a method for RL training of LoRA adapters for specific agentic harnesses. On the Qwen3.8-27B model, it increased the Terminal-Bench 2.1 score from 39% to 55% and reduced token consumption from 2.9K to 1.5K per task. Weights, adapters, and a reproduction recipe are available.
Hugging Face has published a method for training LoRA adapters specific to a given agentic harness using reinforcement learning (RL). Instead of training a model for general performance, the base model weights are frozen, and a small adapter is trained on top directly inside the target harness (Pi, OpenCode, or Hermes Agent), teaching the model to tune its behavior precisely to that tool and type of tasks.
On the Qwen3.8-27B model, the company's data demonstrates an improvement in the Terminal-Bench 2.1 benchmark score from 39% to 55% pass@1 in the Pi harness, while average token consumption per task dropped from 2.9 thousand to 1.5 thousand. According to the company, the improvement is due to the model "thinking" less, calling fewer tools, and succeeding more often on the first attempt. The company subsequently merged three adapters trained this way (one for each harness) into a single "super-adapter," which works across all three harnesses without modification.
The training data was created by combining existing RL datasets (SWE-smith, SWE-Gym) with newly generated tasks from the Tasksmith tool, derived from merged pull requests in real repositories. The entire process ran on the Hugging Face Hub in the Jobs tool, using the TRL library's AsyncGRPOTrainer, OpenEnv, and Harbor. The company has published the weights, adapters, GGUF versions, the set of RL tasks, and the complete reproduction recipe, so the approach can be applied to your own combination of model, harness, and tasks.
The source article contains further technical details on the pipeline, training setup, and open questions that are not available in the excerpt provided.
Why it matters
The method allows an existing model to be adapted to a specific agentic tool without needing to train a new model from scratch — according to the company, a small adapter on top of frozen weights is enough. This can reduce the cost of running agents (fewer tokens per task) and increase task success rate, provided a team verifies the method and trains its own version on its own task distribution. The publication of the weights, data, and recipe also enables independent reproduction and verification of the reported results.
Two audiences, two different impacts
What this means
For individuals
Developers working with agentic harnesses (Pi, OpenCode, Hermes Agent) get a ready-made, reproducible method for using an RL-trained LoRA adapter to increase both accuracy and speed of the model for a specific tool, without changing the base weights.
For a business
Companies running AI agents can use this method to reduce token consumption and increase task success rate, but the impact will only materialize after their own implementation and training of the adapter on their own task distribution.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.