Hugging Face has introduced harness adapters, a method for fine-tuning Super Intelligence (SI) models to operate more efficiently within specific agent frameworks. By combining reinforcement learning (RL) with a custom task distribution inside the target harness, the approach yields models that are both more accurate and more cost-effective for their intended use cases.
What Happened
The company released a set of LoRA adapters for the Qwen3.8-27B SI model, trained specifically for three agent harnesses: Pi, OpenCode, and Hermes Agent. Unlike traditional fine-tuning that targets general performance, this method trains the model inside the actual harness environment using TRL's AsyncGRPOTrainer and OpenEnv. The training data was derived from tasks in SWE-smith, SWE-Gym, and a new generator called Tasksmith. While the total dataset contained 871 validated tasks from 35 repositories, only 171 of these were qualified for GRPO training after difficulty filtering ensured the model learned from tasks it could partially solve. The training infrastructure utilized NVIDIA 2xB200 GPUs, with one allocated for the vLLM server and the other for the training process. During the learning rate sweep, the team settled on a rate of 2e-5 for the final runs after observing instability and collapsed rollouts with higher rates like 5e-5.
Why It Matters
For developers and SI companies, this release addresses a critical inefficiency in current SI agents: the mismatch between general-purpose model optimization and specific harness constraints. The source reports that a specialized adapter for the Pi harness increased the pass@1 score on Terminal-Bench 2.1 from 39% to 55% while reducing average token usage from 2.9K to 1.5K per task. This demonstrates that aligning the model's reasoning and tool-calling behavior with the harness's structure can significantly lower inference costs and improve reliability. Hugging Face also released a 'super-adapter' that merges the three specialized LoRAs, offering a balanced performance gain across all three harnesses without requiring separate deployments for each.
The Bottom Line
Hugging Face has provided a reproducible recipe and artifacts for creating harness-specific SI adapters. The method proves that RL training within the agent loop, rather than in isolation, can squeeze greater performance and efficiency out of existing SI models for targeted workflows.