labs-OO-Agents

Docs  ·  Quick Start  ·  Notebook Tutorials  ·  Examples  ·  Paper  ·  Blog

NVIDIA-labs Object Oriented Agents (NOOA) is a model-agnostic Python framework designed to support reliable AI agent development. Many agent frameworks represent prompts, tools, callbacks, and workflows as separate abstractions. NOOA offers an alternative object-oriented interface that brings these concepts together in a Python class. NOOA lets developers express an agent’s state, capabilities, prompts, and typed interfaces through a single Python class:

from nooa import Agent

# The agent is a Python object.
class SupportAgent(Agent):
    """You are a support agent."""

    # State lives on the object. Fields are typed.
    order_db: OrderDB

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

    # Agentic method: the runtime hands this to an LLM.
    async def triage(self, message: str, order: Order) -> Ticket:
        """Create a typed support ticket."""
        ...

What's happening here:

• Agents are Python objects. Fields are state, methods are capabilities, docstrings are prompts, type annotations are contracts.
• ... bodies are LLM-driven. A method with ... becomes an agentic loop; a real body stays deterministic Python.
• Generation commits one result. Deterministic generators are supported, but generated streams need a separate stream contract and are rejected clearly.
• Code as action. The model acts by writing Python in a Jupyter-style REPL with access to self, imports, and helpers — Python methods and type annotations supply the callable interfaces, reducing the need to write separate tool-schema definitions.
• Pythonic and agent-ready. Typed I/O with auto-retry, live-object arguments passed by reference, and model-callable context and event APIs — designed around agent-oriented Python workflows.

This design supports familiar Python testing, tracing, refactoring, and version-control workflows — just like the rest of your software. Read the paper for the design principles and evaluation results.

Want to see how the pieces compose? Take the 10-minute tour, from one thinking method through tools, typed contracts, deterministic orchestration, and object composition.

Installation

Add the core framework to a new (or existing) Python project with uv:

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

uv add nooa

Or with pip: pip install nooa.

Windows: use WSL 2. Native Windows is not supported: import nooa and the nooa command stop with an error that points to WSL 2. Setting NOOA_SKIP_PLATFORM_CHECK=1 skips that check. The pure-Python core (agents, PredictStrategy, in-process CodeAct, in-memory storage) and nooa connect are tested on native Windows. Other features may work but are unsupported; those that need POSIX (shell tools, the CodeAct sandbox, file-backed SQLite storage, monitor()) stop with a PosixOnlyFeatureError.

The CLI, ACP, memory, and benchmark packages are separate distributions. Install them by name, or pull them in as extras of the core package:

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-bench               # or: uv add "nooa[bench]"

uv add "nooa[cli,memory]"       # several at once

Package	Extra	What it adds
nooa-cli	nooa[cli]	the nooa command, trace-viewer command, eval runner
nooa-acp	nooa[acp]	coding agent for Agent Client Protocol hosts such as Zed — setup
nooa-memory	nooa[memory]	long-term memory subsystem (MemoryManager)
nooa-bench	nooa[bench]	BenchAgent and the Harbor benchmark runner

eval_pipeline is not published to PyPI — install it from the repo:

uv add "eval_pipeline @ git+https://github.com/NVIDIA-NeMo/labs-OO-Agents.git@main#subdirectory=util/eval_pipeline"

# latest development state
uv add "nooa @ git+https://github.com/NVIDIA-NeMo/labs-OO-Agents.git@main"

# pinned to a release tag
uv add "nooa @ git+https://github.com/NVIDIA-NeMo/[email protected]"

Quick Start

⚠️ Before Starting: safety note

NOOA is research software, and agents can be configured to execute LLM-generated code. We welcome contributions and fixes, but expect rough edges. LLM-generated code may take dangerous or unwanted actions, including sending private data to uncontrolled locations, deleting files, or modifying its environments. Ensure you run NOOA agents in a sandboxed environment isolated from your primary filesystem, such as NVIDIA OpenShell.

NOOA validates generated code (AST checks) and applies module deny-lists before execution. These are defense-in-depth guardrails, not a containment boundary. They exist to keep generated code from freezing the event loop and to catch common mistakes early — not to stop code that is actively trying to escape. A static checker over Python cannot provide that guarantee: open() gives arbitrary file access, importlib can load modules straight from a path, and reflection reaches the rest. The containment boundary is OS-level isolation — always run agents that execute generated code inside a sandbox such as a container, VM, or NVIDIA OpenShell. Do not rely on the in-process validators alone.

1. Connect a model

Install the CLI and run the guided setup:

uv add "nooa[cli]"
uv run nooa connect

Choose your provider or endpoint, API key, exact model, and an alias such as my-model. Connect checks the configured route and saves the alias in NOOA's user configuration. The checks send real requests and can cost money; the wizard asks before running them and before saving. Keep keys in environment variables or NOOA's secrets file, not in model configuration or committed code.

Use the saved alias in Python:

from nooa.unifiedllm import get_llm_client

llm = get_llm_client("my-model")

See Connect setup for providers, editing, and diagnostic stages. For setup without inference calls, use offline model configuration. --no-probe alone may still fetch discovery/catalogue metadata; the offline guide explains how to avoid that too. Agent-driven stages do not ask for spending approval; authorize them before use.

You can also construct clients directly using LiteLLM-supported model names, with your endpoint and credentials. See manual configuration.

2. Your first agent

Agents are Python objects. Methods with ... bodies are generation methods — implemented at runtime by an LLM-driven strategy. The signature defines the contract; the docstring is the prompt.

import asyncio

from nooa import Agent


class FeedbackAgent(Agent, llm=llm):
    """You are an agent specializing in analyzing customer feedback."""

    async def analyze_feedback(self, text: str) -> str:
        """Analyze customer feedback for sentiment and key topics in one sentence."""
        ...


async def main():
    try:
        agent = FeedbackAgent()
        result = await agent.analyze_feedback("Great product, but shipping was slow")
        print(result)
    finally:
        await llm.aclose()


asyncio.run(main())

Save the model-selection and agent snippets together as first_agent.py, then run uv run python first_agent.py in your project with approved inference allowance. After your final call, await llm.aclose() releases client resources.

The checked-in first example is a separate source-checkout example: it chooses its own client from environment credentials and does not automatically use your my-model alias.

Rename analyze_feedback to analyze_feedback_briefly and the output changes — your method name, parameters, and docstring are the prompt.

Prefer a guided notebook path? Start with the notebook tutorials, which walk through the same ideas in Colab-friendly steps, with more notebooks planned.

Ready to run something specific? Use the examples catalog to find quickstarts for structured output, tools, strategies, tracing, context blocks, MCP, and more.

3. See what your agent is doing

Every LLM call, code execution, and method invocation is traced by default — orchestrators, generation methods, and helpers, with parent-child spans preserved. If you installed the CLI and viewer dependencies, start the trace viewer and open the run in your browser:

uv add "nooa[cli,viewer]"    # includes viewer runtime dependencies
uv run nooa start-dev        # trace viewer on http://localhost:5001

If the viewer isn't running, tracing is silently disabled — no configuration needed either way.

4. Give your agent media inputs

Pass Image, Audio, Video, or PDF File objects as method arguments. CodeAct can also use show() to see media created while it runs.

NOOA assumes models are text-only until configured. Set media_inputs, such as [image, pdf], in the model entry. Connect lets you choose formats manually or test them with generated samples. If you enable a format the provider does not accept, requests may fail.

from nooa import Agent, PredictStrategy, strategy
from nooa.media import Image
from nooa.unifiedllm import get_llm_client

# This alias must allow image input and use an endpoint that accepts it.
vision = get_llm_client("vision")

class PhotoReader(Agent, llm=vision):
    @strategy(PredictStrategy())
    async def describe(self, image: Image) -> str:
        """Describe the image briefly. Do not invent details."""
        ...

# In your async application, with permission to send the photo:
# result = await PhotoReader().describe(Image.from_file("photo.png"))
# await vision.aclose()

Load media with from_file, from_bytes(..., media_type=...), or from_url. URL support depends on the provider. File supports PDFs. Formats left out of media_inputs are replaced with text notices when sending a request.

See the media API guide for a complete example, CodeAct show(), audio/video/PDF inputs, and limits.

Learn more

• Documentation — human-oriented reading paths, core concepts, architecture, and safety guidance.
• Framework tour — a concise conceptual showcase of NOOA's core ideas and Python-first design.
• Decision models — route classification, yes/no, and scoring methods to a calibrated decision model such as typesafe/jev-1.13 with DecideStrategy; agents that only make decisions need no chat LLM. Runnable example: 16_decisions.py.
• Notebook tutorials — the primary hands-on path for your first agent, strategy selection, CodeAct's live-object workflow, and composing subagents. More notebooks are planned.
• Examples catalog — runnable quickstarts, advanced mechanics, and complete benchmark systems, indexed by capability and setup requirements.
• Paper — design principles, harness details, capability tests, and SWE-bench Verified / Terminal-Bench 2.0 results.
• Blog post — Six Agent Harness Capabilities for Higher Model Performance.
• AGENTS.md — conventions used inside this repo (helpful when reading the source).

Contributing

For a local editable install, clone the repo and sync the development environment with uv:

git clone https://github.com/NVIDIA-NeMo/labs-OO-Agents.git
cd labs-OO-Agents
uv sync --group dev

This installs the core framework, workspace packages, development tools, the nooa CLI, and the trace viewer runtime in the repo's .venv. Run CLI commands through uv:

uv run nooa --help
uv run nooa start-dev       # trace viewer on http://localhost:5001

Enable pre-commit hooks and run the test/lint suite:

uv run pre-commit install
uv run pytest                # run tests
uv run ruff check            # lint
uv run pyright               # type check

See CONTRIBUTING.md for the full workflow.

Citation

If you use NVIDIA-labs Object Oriented Agents in your research, please cite:

@techreport{nvidia_oo_agents_2026,
  title  = {NVIDIA-labs OO Agents: Native Python Object-Oriented Agents},
  author = {Furgale, Paul and Klingler, Severin and Nolan, James and Staats, Matt and
            Di Lorenzo, Gaia and Martinez Abad, Elisa and Schueler, Christian and
            Dinu, Razvan and Devoto, Alessio and Berard, Pascal and Kaplun, Gal and Sarafian, Elad and
            Roveri, Riccardo and Derczynski, Leon and Silveira Cabral, Ricardo},
  year   = {2026},
}

License

Apache 2.0. See LICENSE and THIRD_PARTY_NOTICES.md.