Hugging Face released a new open source project called harness adapters to improve the performance of language models. The toolset is designed to help developers fine tune models using reinforcement learning while reducing the amount of tokens needed for training. It provides ready to use Docker images for three specific agent harnesses named Pi, OpenCode, and Hermes Agent.

The team demonstrated the method on the Qwen3.8-27B model, achieving a significant jump in Terminal-Bench scores from 39 percent to 55 percent. The implementation also cut average token usage from 2,900 to 1,500 tokens. The project includes a custom dataset pipeline that combines 59,000 Hub tasks with generated tasks to create a validated set of 278 training examples.

Developers can use the fully reusable recipe and code to reproduce these results on other task and harness combinations. This release offers a practical way to optimise models for agent workloads without starting from scratch. Users building autonomous agents or complex workflows will find these tools useful for maximising efficiency and output quality.