Infini-AI-Lab has released Hermit, an open-source framework designed as a harness adapter for AI agents. Unlike model adapters such as LoRA, this tool adapts existing agents by incorporating instructions, memory, tools, and coordination logic. The base model and harness remain unchanged, allowing the adapter to function across compatible systems.
The framework utilizes an RL-style training loop consisting of a Trainer, a judge, and an Evolver. The Evolver generates new adapter versions based on task performance, which are then stored in a git tree for version control. The system supports Linux-based sandboxed execution using bubblewrap and offers five built-in evolution methods, including GEPA, SkillOpt, MWM, Harness², and Continual Harness.
Performance benchmarks show varied results across different tasks. In OfficeQA, Harness² achieved a 7.6 point gain over reference runs. In CL-Bench database exploration, MWM reached 0.795 accuracy compared to a 0.580 baseline, while in poker, it won 3.08 big blinds per hand. On the ServeLearnBench, MWM scored 45.3 against a 28.7 score for the CH method.
Hermit is available via PyPI and supports various backends, including OpenAI, Anthropic, DeepSeek, and vLLM. The CLI provides tools for training, evaluation, and deployment. Users can export trained adapters as native plugins for platforms like Codex, Claude Code, and Pi, facilitating collaborative agent workflows.
