Research
Hugging Face shows RL-trained harness adapters lifting Qwen3.8-27B on Terminal-Bench 2.1 from 39% to 55% with about half the tokens, and releases the recipe
On November 20, 2026 Hugging Face described training one LoRA adapter per agent harness (Pi, OpenCode, Hermes Agent) for Qwen3.8-27B with reinforcement learning inside the harness, a merged super-adapter, and a released …