NVIDIA NOOA Explained 2026: An AI Agent Is One Python Class - 82.2% on SWE-bench

Agent frameworks had been getting heavier. NVIDIA’s open-source framework NOOA, released on July 27, 2026, pushes back with a single idea. The short version: treating an agent as one Python class, it scores 82.2% on SWE-bench Verified (up from 79.2%) - and does it using half the tokens (about 1.1M vs 2.2M per task).

What does NOOA change?

The philosophy is one sentence: “An agent is a single Python class.” Instead of stacking prompt templates, custom DSLs, or schema-heavy orchestration, NOOA treats an agent as a Python object - methods are capabilities, fields hold state, and type annotations enforce the contracts.

If you’ve written any Python, it clicks fast. Open the code and you find a readable, debuggable program rather than a pile of prompt magic.

Why the numbers matter

In agents, tokens are cost and speed. So “higher accuracy at half the tokens” carries real weight in production.

Metric	NOOA (GPT-5.5)	Previous best
SWE-bench Verified accuracy	82.2%	79.2%
Tokens per task	~1.1M	~2.2M

That’s a 3-point accuracy gain on the same benchmark while cutting tokens in half. For related model news in the same vein, see our take on Meta Muse and Spark.

What are the six harness capabilities?

NOOA frames its features as an “agent harness.” In practice, four pillars carry it.

• Code as Action — the model runs Python directly to act, rather than faking tool calls through JSON.
• Programmable Loop Engineering — developers control the iteration loop in code.
• Explicit Object State — state lives in typed objects, keeping context management clean.
• Model-callable APIs — the model can inspect and manage its own context and event history.

The direction is clear: make an agent a reproducible Python program. If you’re curious how tools attach to a chat workflow, GitHub MCP and Skills vs MCP sit in the same territory.

How do you start?

It’s on GitHub as NVIDIA-labs’ OO Agents. You just need a Python environment, and the model-agnostic design lets you plug in GPT-family or other models. The flow looks like this.

• Define the agent as a Python class and add capabilities as methods
• Manage state in object fields and control loops in code
• Keep memory in a lightweight store (such as SQLite)

NVIDIA contributed it to the Open Secure AI Alliance it founded, signaling a continued open-source, open-weight direction - so it’s easier to touch than closed frameworks. How much friction it removes in real production, though, you only learn by running it.

Frequently Asked Questions

Q. How is it different from existing agent frameworks? Instead of prompt templates and DSLs, an agent is a single Python class - methods are capabilities, fields are state - which makes it readable and debuggable.

Q. Does it really perform better? On SWE-bench Verified with GPT-5.5 it hit 82.2%, past the previous 79.2%, at about 1.1M tokens per task - roughly half.

Q. Where can I use it? It’s open-sourced on GitHub (NVIDIA-labs OO Agents). The model-agnostic design works with many models, and it was contributed to the Open Secure AI Alliance.