Hermit framework

Hermit is an open-source framework from Infini-AI-Lab for training, evaluating and deploying harness adapters. A harness adapter is a trainable external harness that adapts an existing agent to new scenarios while the original model and harness stay unchanged. The latest release listed is v0.2.0, and the project is licensed under Apache 2.0.

What It Is

Hermit provides a single RL-style training loop for harness adapters. Users bring their tasks, an acting agent, an evaluator and model access, then choose a built-in evolution method or write their own Evolver. An adapter can contain instructions, skills, memory, tools, executable code and agent coordination logic, and it can be an entire repository implementing an external harness.

How the Training Loop Works

The Trainer runs a batch of tasks through the acting agent with the current adapter and scores them with a judge. It then passes trajectories and feedback to an Evolver, which proposes a new adapter. The Trainer checks the proposal against the adapter specification and commits it as the next version. Adapters are stored as files in a git tree, so each version can be inspected, rolled back or shipped.

Methods and Recipes

Supported evolution methods include GEPA, SkillOpt, MetaWorld Model, Harness², Continual Harness, ReHarness, Meta-Harness Adapter, ACE, AHE Adapter, RRSI Adapter and EvoSafeHarness. Runnable recipes cover OfficeQA, CL-Bench and ServeLearnBench. A Lean collaboration example coordinates five Codex agents through a shared adapter to prove a theorem together.

Setup and Deployment

Hermit requires Python 3.12+ and Git. It can be installed from source or from PyPI as hermit-harness, and it exposes a hermit CLI with commands for check, train, eval, deploy and convert. Trained adapters can be deployed as workspaces for Codex, Claude Code or Pi, or converted into native plugins for compatible hosts. Models are referenced by backend, including Fireworks, OpenAI, Anthropic, DeepSeek, vLLM, SGLang and Ollama. The README also describes plan-backed recipes that use a Codex/ChatGPT subscription for acting and learning.