ForkFree

AI agents and automation

reef

by Human-Agent-Society

Reef is a Python framework that serves an AI agent's requests, records how people respond, and uses that feedback to improve the agent over time. It can either retrain the model's weights or tune the agent's prompts, rules and skills, publishing each accepted update as a new version.

Proof level: ReadWorth a lookApache-2.0

Get it on GitHubProject websiteCompare

No demo yet

Proof level: Read. We read its README and licence and ran our automatic checks. We have not run it yet. So this page is based on its README and our automatic checks.

The catch

Worth a lookEstablished, with a few things to check (listed below).

Check this first: The README says the project is being refactored, so some recipes and examples may fail on the main branch.

Things to check

  • The README says the project is being refactored, so some recipes and examples may fail on the main branch
  • Weight training needs a supported GPU stack
  • Young project (open source since 2026) and still changing quickly
  • Harness optimisation needs a model endpoint, which may mean paid API costs

Can you use reef commercially?

Free to use, change and sell, including in paid products. Keep the copyright notice.

Not legal advice. Read the project's licence file before building a product on it.

What can you do with reef?

  • Serve a language model as an inference server and record the interactions
  • Optimise an agent's prompts, rules and skills using example tasks and an evaluator, with no training GPUs
  • Train model weights from live feedback using Slime and SGLang, with versioned releases

How do you get reef running?

Install with uv and pip (`uv pip install reef-infra`, or from a source checkout), with git-lfs required, then start it with the `reef serve` command.

Follow the official instructions in the project README.

Make money with it

A freelancer could offer agent-tuning services, improving a client's existing AI agent by optimising its prompts and skills against the client's own tasks.

An idea, not financial advice. Check the licence before charging for anything built on it.

Ways a business could use it

  • Tune the prompts and rules of a customer-support or workflow agent against real tasks, so it improves without manual rewriting
  • Train a smaller, self-hosted model on your own feedback data to cut per-call costs from paid model APIs

Who's talking about it

Checked 12 Nov 2026. Summary drafted by AI from the project's README. Proof level 1 of 4, Read: we read its readme and licence and ran our automatic checks. we have not run it yet.

Similar tools in AI agents and automation

craft-anything storytold
New and promising

Craft Anything is a desktop app and command-line tool that studies how a website, app or game works, then uses a coding agent to build your own version and check it against the original. It is meant for people who want to recreate or learn from existing software without writing it all by hand.

Proof level: Read 38.9k stars +2044 a day Apache-2.0
dsh-Go ccch1mneyyy
New and promising

DSH Go is an unofficial Android app that runs the DeepSeek Harness coding agent on your phone, or lets you control a copy running on your computer or server. It gives you a chat interface for talking to AI models, approving tool actions and sharing conversations.

Proof level: Read 28.6k stars +1280 a day AGPL-3.0
orca stablyai
Worth a look

Orca is a desktop app, with a phone companion, for running several AI coding agents side by side, each in its own git worktree. It lets you track, review and steer their work from one place.

Proof level: Read 130.3k stars +936 a day MIT