NVIDIA open-sources NOOA: an AI agent is a Python class — and this changes everything in agent development

Announced on July 27, 2026 as a contribution to the Open Secure AI Alliance, the NOOA framework collapses prompts, tools, state, and callbacks into a single Python object — and hits 82.2% on SWE-bench Verified

PrezenceAI Editorial · Published on 08/08/2026

🤖

This article was produced by PrezenceAI with the assistance of artificial intelligence. The main information was extracted from sources cited at the end of this text. The analysis, synthesis, and writing are the sole editorial responsibility of PrezenceAI.

🎧 Listen to this article

0:00 / --:--

On July 27, 2026, NVIDIA Labs announced NOOA (NVIDIA Object-Oriented Agents) — an open-source Python framework for building AI agents. The premise is radical in its simplicity: an agent is a Python class.

Methods are the actions the model can execute. Fields are the agent's state. Docstrings are the prompts. Type annotations are contracts that the runtime enforces. A method with a ... body is completed at runtime by an LLM loop. A method with normal Python code remains deterministic.

Why this matters

Agent development today is fragmented across prompt templates, tool schemas, callback code, and workflow graphs — four different abstractions to represent a single entity. NOOA collapses everything into one class.

The practical consequence: agent behavior can be tested, tracked, refactored, and versioned like normal software. Developers use the same tools they already know — pytest, ruff, pyright — instead of the proprietary infrastructure of each framework.

The benchmarks

NVIDIA reports: 82.2% on SWE-bench Verified, 86.8% on CyberGym L1 (vulnerability rediscovery, with GPT-5.5 and blocked network access), and 85.1% average on ARC-AGI-3. The framework is model-agnostic — it works with any LLM via LiteLLM, including hosted APIs, Ollama, and vLLM endpoints.

The security warning that NVIDIA does not hide

The repository is direct about the risk: NOOA can be configured to execute LLM-generated Python, which can transmit private data, delete files, or modify the environment. AST checks and module deny-lists are defense-in-depth controls — not a containment boundary. The containment boundary is OS-level isolation: container, VM, or NVIDIA's OpenShell.

For regulated production: await stable release. For research teams, AI startups, and enterprise platforms in pilot: available now as a research preview. Installation: pip install nooa (v0.0.8, released Jul 30, 2026, Python 3.12–3.13).

The larger context

NOOA is NVIDIA's main contribution to the Open Secure AI Alliance, a 37-member alliance formed by the company to build and share open-source and open-weight tools for AI development. The strategic signal is clear: NVIDIA wants to be the operating system of agent development, not just the hardware provider.