NVIDIA paper turns AI agents into Python objects

NVIDIA researchers turned the AI agent into a plain Python object in a new paper called NOOA, with Paul Furgale leading 14 co-authors.

For teams building agents, the bet is that ordinary testing and debugging tools should work on agent behavior too. NVIDIA backs it with results across three benchmarks: SWE-bench Verified, Terminal-Bench 2.0, and ARC-AGI-3.

NOOA maps the object model directly onto agent design. A method is an action the model can take. A field holds state.

Docstrings work as the prompts that steer each action, and type annotations become contracts the model has to satisfy before a call goes through.

The design also combines typed inputs and outputs, pass-by-reference access to live objects, and code as the action space itself.

Agents built this way control their own loop logic instead of following a fixed graph. They hold state in explicit fields, and they can call harness APIs directly to pull context or fire events.

I like this more than I expected to. A NOOA agent is a Python class with typed methods, which makes it testable with pytest instead of a mocked graph runtime. The benchmark scores are secondary. What's worth watching is whether other agent frameworks quietly borrow the object model, or keep betting orchestration needs a layer of its own.