Skip to content

Hermit v0.1.0

Latest
Latest

Choose a tag to compare

View all tags
@haizhongzheng haizhongzheng released this 26 Oct 01:17
· 20 commits to main since this release
v0.1.0
deb3856
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

First release of Hermit, a framework for harness adapter evolution, continual learning, and recursive self-improvement.

A harness adapter is a trainable external harness that adapts an existing agent to new tasks while keeping its original model and harness unchanged.

Highlights

  • Trainable external harnesses. Learn instructions, skills, memory, tools, and coordination logic from task feedback, within an explicit adapter specification.
  • A unified training framework. Connect your tasks, agent, evaluator, and Evolver through a shared loop for execution, feedback, adapter updates, evaluation, and versioning.
  • Multiple evolution methods. GEPA, SkillOpt, MetaWorld Model (MWM), Harness², and Continual Harness.
  • Customizable learning workflows. Use the command line or implement your own Evolver and training loop.
  • Deployable adapters. Generate workspaces for compatible agent hosts, including Codex and Claude Code, or export supported adapter contents as native plugins.

Recipes and Examples

  • OfficeQA: GEPA, SkillOpt, Harness², and Continual Harness, with documented evaluation protocols and results.
  • CL-Bench and ServeLearnBench: MWM and prompt-only recipes for learning from task streams.
  • Lean Collaboration: a finished harness adapter that enables five Codex agents to coordinate through shared plans, messages, and verified lemmas.
  • Offline Quick Start: a complete training, evaluation, and deployment walkthrough without model access or a GPU.

See the example collection and recipe catalog.

Install

Requires Python 3.11+ and Git.

git clone --branch v0.1.0 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

GEPA, SkillOpt, and Harness² require their corresponding requirements-*.txt dependencies. See the installation guide and each recipe for additional requirements.

Status

Early release — APIs and configuration formats may evolve during the 0.x series. Feedback, new Evolvers, adapter designs, and task integrations are welcome.

Report an issue · Full changelog

Assets 2
Loading