Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost
Summary
- Hugging Face introduced harness adapters, using RL to train LoRA adapters that optimize base models for specific agent harnesses and task distributions.
- Training Qwen3.8-27B inside the pi harness with a custom LoRA adapter improved its Terminal-Bench 2.1 score from 39% to 55% while reducing average token usage.
- The team trained separate LoRA adapters for Qwen3.8-27B across three agent harnesses: pi, Hermes Agent, and OpenCode using TRL's AsyncGRPOTrainer and OpenEnv.
- Individual harness adapters were merged into a single super-adapter that runs unchanged across all three evaluated agent harnesses.
Why it matters
Harness adapters bridge the gap between general-purpose base models and specific agent execution frameworks, significantly improving both task accuracy and token efficiency without modifying base weights.