nvidia-nemo/labs-OO-Agents ? reverse-engineered prompt

Build me a Python library for making AI agents as normal Python objects. I want to define an agent as a class where typed fields hold state, regular methods do normal deterministic work, and special async methods with empty bodies become LLM powered steps. The library should support typed inputs and outputs, automatic retry when the model gives the wrong shape, and a simple way to pass context, events, and helper methods through the agent object itself.

Please include a small set of clear examples and quick start docs, plus tests that show how to define an agent, call model driven methods, and use a local or hosted model through one unified interface. It should feel Pythonic and easy to read, with good error messages and a clean public API. If you need to check current docs for model providers or best practices, look them up online.