NOOA (NVIDIA Object Oriented Agents)

// Overview

NVIDIA-labs Object Oriented Agents (NOOA) is a model-agnostic Python framework for building agents, published by NVIDIA under Apache-2.0 with an accompanying arXiv paper. Its argument is that most agent frameworks scatter prompts, tools, callbacks and workflows across separate abstractions, and that a plain Python class can hold all of them at once.

So in NOOA an agent is a class. Its fields are state and are typed. Its docstring is the system prompt. Ordinary methods with real bodies stay deterministic Python. A method whose body is `...` becomes an agentic loop: the runtime hands it to an LLM, and the type annotations act as the contract for what comes back, with automatic retry on typed I/O. The model acts by writing Python in a Jupyter-style REPL that has access to `self`, imports and helpers — so the methods and type annotations you already wrote supply the callable interfaces, and separate tool-schema definitions are largely unnecessary.

The practical consequence is that an agent is an ordinary software object: you can test it, trace it, refactor it and diff it with the tools you already use. The core `nooa` package installs on its own; a CLI with a trace viewer and eval runner, an Agent Client Protocol coding agent for hosts such as Zed, a long-term memory subsystem and a benchmark runner ship as separate distributions or extras. The repository also carries a documented ten-minute tour, notebook tutorials and examples, and an evaluation pipeline installable straight from the repo.

// What it does

• Agents are Python classes: typed fields are state, methods are capabilities, docstrings are prompts, annotations are contracts
• A method body of `...` becomes an LLM-driven agentic loop; a real body stays deterministic Python in the same class
• Code-as-action — the model writes Python in a Jupyter-style REPL with access to `self`, imports and helpers, so tool schemas are mostly redundant
• Typed I/O with automatic retry, live objects passed by reference, and model-callable context and event APIs
• Optional sub-packages: `nooa-cli` (command, trace viewer, eval runner), `nooa-acp` (Agent Client Protocol coding agent), `nooa-memory`, `nooa-bench`
• Model-agnostic, Apache-2.0, with a published paper on the design and evaluation

// Getting started

NOOA is a normal Python package. The commands below are from the repository's installation section; the class sketch is the README's own example.

uv init my-agent-project
cd my-agent-project

uv add nooa
# or: pip install nooa

from nooa import Agent

class SupportAgent(Agent):
    """You are a support agent."""

    order_db: OrderDB

    def is_refund_eligible(self, order: Order) -> bool:
        return order.delivered and order.days_since_delivery <= 30

    async def triage(self, message: str, order: Order) -> Ticket:
        """Create a typed support ticket."""
        ...

uv add nooa-cli                 # or: uv add "nooa[cli]"
uv add nooa-acp                 # or: uv add "nooa[acp]"
uv add nooa-memory              # or: uv add "nooa[memory]"
uv add "nooa[cli,memory]"       # several at once

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Building agents you can unit-test, refactor and code-review like the rest of your Python
• Mixing deterministic logic and model-driven steps inside one object instead of splitting them across a graph
• Getting typed, retried outputs out of an agent without hand-writing tool schemas
• Running a coding agent inside an Agent Client Protocol host such as Zed via `nooa-acp`

// How NOOA (NVIDIA Object Oriented Agents) compares

NOOA (NVIDIA Object Oriented Agents) alongside other open-source agent frameworks & builders tools AI/TLDR tracks, ranked by GitHub stars.