Hugging Face introduces harness adapters to tune models for agent harnesses
Hugging Face released a harness-adapter method that trains LoRA adapters with RL inside a specific agent harness, lifting Qwen3.8-27B's Terminal-Bench 2.1 pass rate from 39% to 55% while cutting average tokens per task from 2.9K to 1.5K. The approach spans Pi, OpenCode and Hermes Agent, with a reusable recipe published.