NVIDIA AI Releases NOOA: An Object-Oriented Python Framework
NVIDIA Labs introduces NOOA, an Apache 2.0 framework that collapses complex AI agent development into a single Python class.
Stock photo for illustration only, not from the actual event
- NVIDIA Labs releases NOOA, a model-agnostic Python framework for AI agents
- Collapses prompts, tool schemas, and workflows into one single class
- Supports pluggable models via LiteLLM and installs simply via pip install nooa
- Achieves 82.2% on SWE-bench Verified when paired with GPT-5.5
NVIDIA Labs has rolled out NOOA (NVIDIA Object-Oriented Agents), a model-agnostic Python framework designed for building agents. Traditional agent workflows typically scatter source code across prompt templates, tool schemas, callbacks, and graph structures, but this new framework consolidates all those moving parts into a single class.
Under the hood, methods represent actions the model can take, fields manage state, and docstrings act as prompts. Type annotations serve as contracts enforced strictly by the runtime. A method body left empty becomes an agentic method completed dynamically by an LLM loop, whereas a standard method body remains deterministic Python that the model can invoke as a tool.
The research team highlights six model-facing concepts combined together for the first time: typed input/output, pass by reference over live objects, code as action, programmable loop engineering, explicit object state, and model-callable harness APIs. NVIDIA benchmarked fourteen other frameworks and harnesses, including LangGraph, Google ADK, PydanticAI, and smolagents, reporting only partial feature coverage elsewhere.
Stock photo for illustration only, not from the actual event
NVIDIA's release of NOOA represents a major paradigm shift in simplifying AI agent engineering. By blending agent workflows with familiar Object-Oriented Programming (OOP) principles, developers can drastically reduce token overhead and optimize context window management during complex execution loops.
Efficiency stands out as a core advantage, delivering an 82.2% benchmark score using roughly 1.1 million tokens and about 28 model calls per task, compared to competing tools that consume 2.2 million tokens and 66 calls. Furthermore, the framework enforces validated termination, requiring a typed TaskResult containing verifiable evidence rather than stopping blindly when a tool call ends.
The framework operates under the Apache 2.0 license, installs via pip install nooa (version v0.0.8 released on July 30, 2026), and requires Python 3.12 through 3.13. PyPI classifies the package as alpha, while NVIDIA labels it a research preview, noting that its AST checks and deny-lists function as defense-in-depth guardrails within OS-level isolation.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment