NVIDIA NOOA: build an AI agent as one Python class

Most agent tutorials scatter prompts, tool schemas, callbacks, and workflow graphs across several files. NOOA (NVIDIA Object-Oriented Agents) collapses that surface into one Python class: methods are actions the model may take, fields are agent state, docstrings are prompts, and type annotations are contracts the runtime enforces. A method whose body is ... is completed at runtime by an LLM-driven loop; a method with a normal body stays deterministic Python the model can call as a tool.

That mapping is a strong after-work intro to agents if you already know — or are learning — classes. It is also a research preview / alpha: execute generated code only inside OS-level isolation (container, VM, or similar). Do not treat it as a production install for an office laptop tonight.

Sources: MarkTechPost — NVIDIA AI Releases NOOA (7 August 2026); GitHub — NVIDIA-NeMo/labs-OO-Agents.

One class, four contracts

MarkTechPost summarises the design NVIDIA Labs open-sourced:

Python surface	Agent meaning
Method	Action the model can take
Field	Mutable / readable agent state
Docstring	Prompt text for that surface
Type hints	Contracts enforced by the runtime
Body ...	Agentic method — LLM loop at runtime
Normal body	Deterministic tool the model may call

Two strategies ship in the coverage MarkTechPost cites: PredictStrategy (a single typed LLM call with a local retry loop on validation failure) and CodeActStrategy (an iterative Python REPL where the model calls execute_python(...) until it submits a typed return_result(...)).

You do not need to memorise every strategy name after dinner. The useful habit is: one object, explicit state, typed boundaries. That is ordinary OOP — then the LLM fills the ... methods.

Why this helps evening learners

After work you might have 45 minutes. Graph-heavy agent frameworks ask you to learn a second vocabulary (nodes, edges, callbacks) before you can print a result. NOOA's bet is that if you can write a class, you can sketch an agent:

# Illustrative shape only — see the GitHub repo for current APIs.
class NightlySummariser:
    """Summarise a short work note for tomorrow's standup."""

    notes: list[str]

    def add_note(self, text: str) -> None:
        self.notes.append(text)

    def summarise(self) -> str:
        ...

add_note is ordinary Python. summarise with ... is the agentic hole. Type hints and the docstring become the contract and the prompt. You can version-control, review, and refactor the class like any other module.

MarkTechPost also highlights pass by reference over live objects: large data can stay in the REPL while the model sees a bounded preview (type, length, head/tail sample). That is an advanced detail — the after-work lesson is still "keep big objects in memory; do not paste them into the prompt."

Alpha means isolate execution

NOOA is described as Apache 2.0 and pip-installable (pip install nooa in MarkTechPost's write-up; PyPI classified alpha; Python 3.12–3.13). NVIDIA's own framing, as reported there: AST checks and module deny-lists are defence-in-depth, not a containment boundary. The containment boundary is a container, a VM, or NVIDIA OpenShell. Models plug in through LiteLLM (hosted APIs, Ollama, vLLM, and similar).

So the responsible evening experiment is: read the repo, run demos only where generated code cannot touch your personal files or office credentials, and treat every ... method as "someone else's code ran on my machine."

Benchmark numbers appear in the MarkTechPost article (NVIDIA-reported scores on public agent benchmarks). This Hub post does not invent SoloQuant metrics and does not ask you to chase a leaderboard after dinner. If you care about numbers, read the cited article and the paper/blog it links; then come back to classes and types.

What SoloQuant is (and is not)

SoloQuant Academy is short, hands-on, after-work Python for employed people who want to stay competitive in the job they already have. It is not a career-switch bootcamp and not a get-rich trading pitch.

Before you chase agent frameworks, get comfortable with classes, type hints, and a real .py file you can re-run. The live path is Intro to Python Programming. Check the course page for the current enrollment window.

NOOA's value for after-work learners is the metaphor: an agent is still software. Methods, fields, docs, and types are the same tools you use for ordinary Python — then an LLM fills the ellipsis under a contract you wrote.