Hugging Face shows harness-specific adapters can make coding agents leaner and more accurate
Hugging Face has published a reproducible reinforcement-learning recipe for adapting Qwen3.8-27B to specific coding-agent harnesses. Specialist LoRA adapters improved benchmark accuracy while cutting tool calls and generated tokens, and a merged super-adapter transferred gains across three harnesses. The results highlight deployment context as a trainable part of agent performance.