Hermit — Train Your Harness Adapter

§Supported Harness Evolution Methods

Use an existing method, or bring your own Evolver to the same training framework.

GEPA SkillOpt MetaWorld Model Harness² Continual Harness

Coming soon: ACE · ReHarness. Compare methods and requirements →

§What Hermit gives you out of the box

Hermit is an open-source framework with one training loop for harness adapters. Bring your tasks and feedback; choose a learning method or write your own. It provides shared infrastructure for:

One Trainer for every method: rollouts, judge feedback, an Evolver step, a checked commit, the next version
Built-in evolution methods and a shared interface for your own Evolver
Adapters as files in a git tree: inspect every version, roll back, ship the final one
Deploy workspaces for Codex, Claude Code or Pi, or export supported contents as a native plugin for compatible hosts
Adapter specification checks and versioned updates, with sandboxing for supported learner programs
Any model backend by reference: Fireworks, OpenAI, Anthropic, DeepSeek, vLLM, SGLang, Ollama

Open source · Apache 2.0

Train an adapter for your agent on your own tasks.

Star on GitHub Quick start

§Why a harness adapter

Need	Agents meet scenarios their harness was not written for: a new domain, a new tool, a stream of tasks whose rules only show in the feedback. They should keep learning in deployment.
Today	Two ways to adapt an agent: edit its harness by hand, which does not scale and does not learn, or fine-tune the model, which is expensive, opaque, and tied to one model.
Why	A harness adapter puts the learnable part outside the agent: an instruction file, skills, tools, a world model of the venue. The base stays frozen, the adapter trains from feedback, and the same adapter moves between compatible harnesses.

§How it works: an RL-style loop for harnesses

Training flow: tasks run through the acting agent and harness adapter, feedback reaches the Evolver, and checked updates become the next adapter version. — **One loop, many methods**The Trainer runs a batch of tasks through the acting agent with the current adapter, scores them with your judge, and hands the trajectories and feedback to an Evolver. The Evolver proposes a new adapter; the Trainer checks it against the template and commits it as the next version.

From the command line

Install Hermit, train an adapter, deploy it, convert it into a plugin.

Or write your own loop

Training pseudocode: initialize, rollouts, optimize, check and commit, finalize.

Implement your update logic in an Evolver. The Trainer manages specification checks, versioning, run records and evaluation support.

§Examples and recipes

A harness adapter for collaboration

Five Codex agents, one Lean theorem. A shared harness adapter connects their plans, messages and verified lemmas, so each agent can build on the others’ work without modifying Codex itself.

Five fixed Codex agents communicate through a shared harness adapter while their verified lemma library and collaborative proof grow over time. — **Shared adapter, coordinated agents**The diagram shows how agents communicate through the adapter; the replay below shows their contributions to one proof.

Explore the Lean collaboration adapter →

Train adapters on your tasks

Start from a runnable recipe, connect your tasks and feedback, and choose an evolution method.

OfficeQA

Optimize document-analysis skills or adapt throughout a task stream with GEPA, SkillOpt, Harness² and Continual Harness.

Explore recipe →

CL-Bench

Build environment knowledge from experience with MWM and prompt-only adapters.

Explore recipe →

ServeLearnBench

Learn a changing policy while serving requests, then evaluate the adapter on held-out tasks.

Explore recipe →

§Get started

Requires Python 3.11+ and Git.

From source (recommended)

Clone the repository to run examples, reproduce recipes and develop your own evolution methods.

git clone https://github.com/Infini-AI-Lab/hermit.git
cd hermit
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
hermit --help

Also available on PyPI

Install the framework CLI and bundled adapter templates in your Python environment:

python -m pip install hermit-harness
hermit --help

The included examples and benchmark recipes use files from the source checkout. See Installation for method-specific dependencies.

Offline Quick Start · train with an API model · run OfficeQA · bring your own task · build a new method