NVIDIA Open-Sources NOOA Framework for Building AI Agents

AI agent performance depends not only on the underlying model but also on the harness that manages prompts, tools, state and execution. NVIDIA Labs’ open-source research preview, NVIDIA Labs Object-Oriented Agents, or NOOA, consolidates those components into a single Python class. Methods define capabilities, fields hold state, docstrings serve as prompts and type annotations enforce contracts, allowing agents to be unit-tested, traced, versioned and refactored using familiar software-development practices.

NVIDIA published NOOA’s framework, memory system, capability tests and evaluation code on July 27, 2026. A general-purpose 253-line agent powered by GPT-5.5 scored 82.2% on SWE-bench Verified, topping the 79.2% public state of the art at submission. It used 29 LLM calls and about 1.1 million tokens per task, versus 66 calls and 2.2 million tokens for a comparison harness that scored 78.2%, demonstrating roughly half the token cost.