OpenPipe/ART ? reverse-engineered prompt

Build me a Python project that lets me train an agent to do multi step tasks with reinforcement learning, kind of like teaching a chatbot by giving it a goal, a reward, and repeated tries. I want the main experience to be easy for someone to run from a notebook or a simple script, with a few working examples like a small game, a search or tool use task, and a basic fine tuning example.

It should support training and evaluating an agent, saving checkpoints, and showing progress over time. If possible, make it work with a few popular open models and keep the setup simple enough that I can try it locally, but also be able to plug in managed infrastructure later. Please include clear docs, a couple of runnable examples, and sensible defaults so I can get something working quickly and then customize the reward and environment.

If you need current library details, look them up online first.