Repolode
Submit
Infini-AI-Lab

Infini-AI-Lab/hermit

A framework for training, evaluating and deploying harness adapters that evolve an external layer around an agent without changing its model or harness.

View on GitHubWebsite
Stars
491
Forks
45
Open issues
11
Last commit
21 hours ago

What is it?

Hermit is a framework for training harness adapters: external, trainable layers that adapt an existing AI agent to new tasks. The agent's model and original harness stay unchanged, while the adapter can contain instructions, skills, memory, tools, executable code and coordination logic that get updated from task feedback.

The project provides a training loop in which an Evolver proposes changes to the adapter, checks them, and commits them as new versions. It currently includes several harness evolution methods, among them GEPA, SkillOpt, MetaWorld Model, Harness² and Continual Harness, along with a command-line interface for checking, training, evaluating, deploying and converting adapters.

Adapters can be deployed into a workspace for supported agent hosts such as Codex, or converted into native plugins for Codex, Claude Code, pi and Hermes. Recipes cover benchmarks such as OfficeQA, CL-Bench and ServeLearnBench, and an offline quick start runs without a model account or GPU.

Key features

  • 01Trains external harness adapters without modifying the original model or harness.
  • 02Supports several harness evolution methods, including GEPA and SkillOpt.
  • 03Provides a CLI for checking, training, evaluating and deploying adapters.
  • 04Converts adapters into native plugins for Codex, Claude Code, pi and Hermes.
  • 05Includes an offline quick start that needs no model account or GPU.

Use cases

  • Adapting a closed-source coding agent to a new task using task feedback.
  • Comparing harness evolution methods on shared benchmark pipelines.
  • Packaging a learned agent adaptation as a plugin for another agent host.

Quick start

  1. 1

    Clone the repository (requires Python 3.11+ and Git)

    git clone https://github.com/Infini-AI-Lab/hermit.git
  2. 2

    Change into the checkout and install Hermit in editable mode in a virtual environment

    python -m pip install -e .
  3. 3

    Run the offline local training walkthrough (no model account, GPU or dataset needed)

    hermit train --config docs/_examples/first_training/run.yaml

Things to know

  • Version 0.1.0 is early-stage, and the project is young with few stars.
  • The offline demo's score gain is constructed and is not a real-model performance result.
  • Some methods need extra dependencies, and sandboxed agent Evolvers require Linux with bubblewrap.

Topics

#ai#agent#llm#continual-learning#harness#plugin#evaluation

FAQ

What is hermit?
A framework for training, evaluating and deploying harness adapters that evolve an external layer around an agent without changing its model or harness. Hermit is a framework for training harness adapters: external, trainable layers that adapt an existing AI agent to new tasks. The agent's model and original harness stay unchanged, while the adapter can contain instructions, skills, memory, tools, executable code and coordination logic that get updated from task feedback.
Is hermit free to use? What license does it have?
hermit is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is hermit still maintained?
Actively maintained. The last commit to hermit was on Nov 6, 2026.
What can I use hermit for?
Adapting a closed-source coding agent to a new task using task feedback. Comparing harness evolution methods on shared benchmark pipelines. Packaging a learned agent adaptation as a plugin for another agent host.
How do I install hermit?
Clone the repository (requires Python 3.11+ and Git): git clone https://github.com/Infini-AI-Lab/hermit.git Change into the checkout and install Hermit in editable mode in a virtual environment: python -m pip install -e . Run the offline local training walkthrough (no model account, GPU or dataset needed): hermit train --config docs/_examples/first_training/run.yaml
Where can I get hermit?
The source code and installation instructions are on GitHub: https://github.com/Infini-AI-Lab/hermit

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects