Infini-AI-Lab Releases Hermit, an Open-Source Harness Adapter Framework for AI Agents
Summary
- Infini-AI-Lab has released Hermit v0.1.0, an open-source Python framework that trains external harness adapters for AI agents while keeping model weights and base harnesses frozen.
- The RL-style training loop uses a Trainer, a judge, and an Evolver to update adapters, which can contain instructions, skills, memory, and tools.
- Hermit ships with five evolution methods including GEPA, SkillOpt, MetaWorld Model (MWM), Harness2, and Continual Harness.
- Self-reported benchmarks using GLM 5.3 Flash show gains, such as MWM reaching 0.795 accuracy on CL-Bench database exploration compared to an in-context baseline of 0.580.
- Trained adapters can be deployed to Codex, Claude Code, and Pi, or exported as native plugins under the Apache-2.0 license.
Why it matters
Hermit offers a modular approach to agent optimization by separating training from the underlying model and harness, allowing developers to improve agent capabilities across closed-source or fixed systems without fine-tuning.