NVIDIA Releases NOOA Framework to Streamline AI Agents

NVIDIA has launched the OO Agents framework, known as NOOA, which unifies prompts, tools, state, and model-driven loops within a single Python class. In this setup, standard Python methods manage deterministic logic, while methods defined with ellipsis bodies are dynamically completed by the AI model at runtime. This design addresses the fragmentation of modern agent development, where tasks are often scattered across mismatched configuration systems.

To execute these generation methods, NOOA offers two distinct strategies. The first, PredictStrategy, handles single structured predictions. The second, CodeActStrategy, provides the model with an iterative Python REPL that has direct access to live object states. By utilizing live objects, the framework avoids the repeated JSON serialization round-trips that typically slow down agent execution.

While this live-object approach eliminates serialization overhead within a single process, it introduces a significant trade-off. NOOA lacks built-in cross-session recovery, making durability, distribution, and replay more difficult than in graph-based runtimes that prioritize checkpointing. Despite these architectural challenges, NVIDIA reports competitive benchmark performance for NOOA on SWE-bench Verified, Terminal-Bench 2.0, and ARC-AGI-3.

For AI practitioners, NOOA represents a shift in how agent runtimes are structured. Instead of relying on rigid graph workflows or complex tool schemas that only define action boundaries, developers can write agents using familiar object-oriented programming. This simplifies the management of loop ownership and state location, though developers must still weigh the benefits of execution speed against the need for robust session durability.