Infini-AI-Lab releases Hermit, an open-source adapter framework for AI agents

Infini-AI-Lab has released Hermit, an open-source adapter framework designed for AI agents. The framework lets developers keep an agent’s model and core program frozen while training an external adapter, allowing agents such as Codex and Claude Code to take on new tasks without fine-tuning. The first release, v0.1.0, went out on Oct. 25, 2026. Hermit uses a reinforcement learning-style training loop, and its adapters can include instructions, skills, memory, tools, and executable code. In the current version, the framework offers five adapter evolution methods, including GEPA, SkillOpt, and MetaWorld Model. The team reported performance gains on benchmarks including OfficeQA and CL-Bench. Carnegie Mellon University postdoctoral researcher Haizhong Zheng announced the release on X on Oct. 27, according to the Techub summary citing MarkTechPost.