Infini-ai lab unveils Hermit, a new open-source framework for AI agent adaptation

Infini-AI Lab has launched Hermit, an open-source harness adapter framework designed to enable AI agents to adapt to new tasks without requiring fine-tuning or forking the harness. The framework, released as version 0.1.0 on October 25, 2026, allows developers to train an external 'harness adapter' while keeping the agent’s model and harness unchanged. CMU postdoc Haizhong Zheng announced the release on X on October 27. Hermit supports multiple evolution methods, including GEPA, SkillOpt, and MWM, with additional methods like ACE and ReHarness planned for future updates. The framework enables adapters to hold instructions, skills, memory, and coordination logic, and can even function as a full repository. Training uses an RL-style loop, with a Trainer evaluating outputs and an Evolver proposing new adapter versions. Results from benchmark tests show significant improvements, with MWM achieving 0.795 accuracy on database tasks and 3.08 big blinds per hand in poker. Developers can install Hermit via PyPI or clone the repository, with support for multiple backends including Fireworks and OpenAI.