Infini-AI-Lab has released Hermit v0.1.0, an open-source harness adapter framework designed to adapt AI agents to new tasks without fine-tuning or forking the base model. Announced by CMU postdoc Haizhong Zheng on October 27, 2026, the Python framework trains an external harness adapter while keeping the underlying model and agent harness completely frozen.
The Apache-2.0 licensed framework utilizes an RL-style training loop featuring a Trainer, a judge, and an Evolver to propose, evaluate, and commit new adapter versions within a git-managed tree. Hermit ships with five built-in evolution methods and supports backends including OpenAI, Anthropic, DeepSeek, vLLM, and Ollama. In self-reported benchmarks using GLM 5.3 Flash, the MetaWorld Model adapter achieved 0.795 accuracy on CL-Bench database exploration compared to a 0.580 in-context baseline, while Harness² reported a 7.6 point gain over reference runs on OfficeQA.
This framework matters because it allows developers to improve and customize closed or brittle agent systems through external plugins without risking disruptions to carefully tuned component interactions.