Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost
Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost
The post describes training LoRA adapters for Qwen3.8-27B with RL inside three agent harnesses (Pi, OpenCode, Hermes Agent), using TRL's AsyncGRPOTrainer, OpenEnv and Harbor, with datasets built from SWE-smith, SWE-Gym and Tasksmith and filtered by validation and a base-model difficulty check. Each specialist adapter beats the base model under its own harness on Terminal-Bench 2.1 and the hard subset of SWE-bench Pro while making 53–72% fewer tool calls, and a merged super-adapter (rank 96, weight ½) improves on the base model under all three harnesses. On 300 AutomationBench tasks run in a different agent loop, the adapters stayed close to the base model, which the authors read as neither clear transfer nor breakage.
Useful for teams weighing harness-specific RL fine-tuning, since it gives a reproducible recipe plus paired benchmark and efficiency numbers, including where gains do and do not transfer.
Source: Hugging Face Blog · huggingface.co