NVIDIA Labs Releases NOOA Framework for Object-Oriented AI Agents
NVIDIA Labs has released NOOA (NVIDIA Object-Oriented Agents), an Apache 2.0 Python framework for building model-agnostic AI agents. It represents an agent as one Python class: methods define available actions, fields hold state, docstrings serve as prompts, and type annotations act as runtime-enforced contracts. Methods containing ... are completed through an LLM-driven loop, while regular method bodies remain deterministic Python code.
NOOA includes PredictStrategy for typed calls with local validation retries and CodeActStrategy, which uses an iterative Python REPL until the model returns a validated result. Its design supports typed inputs and outputs, live-object references, programmable loops, explicit state, and model-callable APIs. Models can be connected through LiteLLM, including hosted services, Ollama, and vLLM endpoints. An optional memory system stores records in SQLite.
NVIDIA reports 82.2% on SWE-bench Verified and 86.8% on CyberGym L1, with about half the tokens used by compared open harnesses. The package requires Python 3.12–3.13, is classified as alpha, and was released as version 0.0.8 on July 30, 2026. Because agents can run generated code, NVIDIA says AST checks and module deny-lists are not a containment boundary; deployment should use a container, virtual machine, or NVIDIA OpenShell.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.