NVIDIA Labs Releases OO Agents: Open-Source Python Framework for Object-Oriented AI Agent Development
Key Takeaways
- ▸NOOA brings agent development into Python's standard class paradigm, making agent construction more intuitive for Python developers
- ▸The object-oriented design integrates state, methods, and prompts into a single abstraction, simplifying agent architecture
- ▸Framework supports typed I/O, auto-retry logic, and live-object arguments, reducing boilerplate code and complexity
Summary
NVIDIA Labs has released NOOA (NVIDIA-labs OO Agents), an open-source, model-agnostic Python framework designed to simplify AI agent development through a native object-oriented interface. Rather than treating prompts, tools, callbacks, and workflows as separate abstractions, NOOA unifies these concepts into a single Python class, allowing developers to define agents much like they would any other Python object.
The framework enables developers to express agent state as typed fields, capabilities as methods, and prompts as docstrings—all within a familiar Python class structure. Methods marked with ellipses ("...") become agentic loops handed to an LLM, while methods with full implementations remain deterministic Python. The runtime executes LLM-generated code in a Jupyter-style REPL with access to the agent's methods and type annotations, reducing the need for separate tool-schema definitions.
NOOA is now available on PyPI and can be installed with pip or uv. The core framework comes with optional extensions including a CLI with trace viewer, a long-term memory subsystem, and benchmarking tools. NVIDIA emphasizes that this is research software and agents should be run in sandboxed environments, with built-in AST validation and module deny-lists as defense-in-depth guardrails.
- Open-source release with model-agnostic design allows use with any LLM provider
