NVIDIA Releases NOOA, an Open-Source Framework That Treats AI Agents as Python Classes

NVIDIA Labs has released NOOA (NVIDIA Object-Oriented Agents), a model-agnostic Python framework for building AI agents. The framework consolidates agent logic—methods, state, and prompts—into a single Python class, simplifying development compared to traditional approaches that scatter logic across prompt templates, tool schemas, callbacks, and workflow graphs. Agentic methods with a body of ... are completed at runtime by an LLM-driven loop, while normal methods maintain deterministic Python behavior. The framework supports two strategies: PredictStrategy for simple LLM calls and CodeActStrategy for iterative Python REPL loops. Testing shows strong results—82.2% on SWE-bench Verified, 86.8% on CyberGym L1, and 85.1% on ARC-AGI-3—using roughly half the tokens of comparable open-source harnesses. NOOA is Apache 2.0, requires Python 3.12-3.13, and is available as a research preview with pluggable models via LiteLLM supporting hosted APIs, Ollama, and vLLM endpoints. Security notes indicate that code execution is possible but the containment boundary is OS-level isolation via containers, VMs, or NVIDIA OpenShell.