Hugging Face 发布 harness adapters：用安全 RL 高效开发 agent
原标题：Harness adapters: Safe RL for efficient agentic development
AI 摘要
Hugging Face 博客发布“harness adapters”方法：为每个 agent harness（pi、opencode、hermes）单独训练小型 LoRA 适配器，基于同一基础模型，用强化学习在可验证任务上训练，并在工具调用前由 safeagents 代理做规则检查以阻止 reward hacking。实验显示适配器在留出评测任务上工具调用更少、准确率保持或提升，训练覆盖 LFM2.5-24B-A2B、LFM3-1.7B 及未做 agent 训练的 TwIL-8B。
正文节选
Small LoRA adapters, trained with reinforcement learning inside pi, opencode and hermes, all served from one model, with every tool call the model makes inspected by safeagents before it runs. Follow the training runs on the live dashboard. - We release several "harness adapters". It is one LoRA adapter per model and harness, trained with RL on verifiable tasks, inside unmodified agent harnesses (pi, opencode and hermes). - Harness adapters can make local agentic development more efficient and a