openpipe/art ? reverse-engineered prompt

Build me an open source Python project for training multi step AI agents with reinforcement learning.

I want a clean framework that lets me define an agent, a task environment, and a reward function, then train with GRPO and related methods. It should support common model families like Qwen, Llama, GPT OSS, and LoRA based fine tuning, with a simple API that makes it easy to plug into an existing Python app.

Please include a few working examples so I can see it in action, like a small game or puzzle task, a document style task, and an agent that uses tools or a search like workflow. Add clear docs, setup instructions, and a notebook or demo path that shows how to run training end to end. If it helps, use current docs online for the latest model and tooling details. Make the project feel production friendly, with tests, sensible defaults, and an easy way to run locally or on managed infrastructure.