NVIDIA launches NOOA: object-oriented Python framework for AI agents

NVIDIA Labs released NOOA (NVIDIA Object-Oriented Agents) as an open-source, Python-agnostic framework for building AI agents. Traditional agent development is currently spread across prompt templates, tool schemas, callback code and workflow graphs; NOOA collapses all of these into a single Python class. Methods represent the actions a model can perform, fields hold agent state, docstrings serve as prompts, and type annotations act as contracts enforced at runtime. A method whose body is `...` becomes a stub method that is completed at execution time by an LLM-driven loop; a method with a regular body remains deterministic Python that the model can read as a tool.

The company reports 82.2% on SWE-bench Verified, 86.8% on CyberGym L1 and an 85.1% average RHAE on ARC-AGI-3, using roughly half the tokens of the open-source toolkits evaluated against them. The numbers come from NVIDIA’s own report rather than an independent audit, and the framework is listed on PyPI as an alpha classified by the company as a research preview. Version 0.0.8 was released on 30 July 2026 and requires Python 3.12-3.13; installation is performed with `pip install nooa` under an Apache 2.0 license.

NOOA agents can execute code generated by the LLM, and NVIDIA states directly that AST checks and module blocklists are guardrails of defense-in-depth, not containment limits. Containment is provided by a container, VM or NVIDIA OpenShell. Models can be connected via LiteLLM, so hosted APIs, Ollama endpoints and vLLM are all supported. The company advises organizations with regulated production workloads to wait for a stable release.

Two execution strategies ship with the framework: PredictStrategy, which makes a single LLM call with a local retry loop on failure and validation; and CodeActStrategy, an iterative Python REPL where the model repeatedly calls `execute_python(...)` until it submits `return_result(...)`, which is then validated against the return-type annotation. The research team notes that capabilities that “reach the model,” in their view, are the first to be introduced: token-level I/O, pass-by-reference over live objects, code-as-action, programmable loop engineering, explicit object state, and APIs of toolkits readable by the model. NVIDIA evaluated fourteen frameworks and toolkits—including LangGraph, Google ADK, PydanticAI, smolagents, Claude Agent SDK, OpenAI Codex and OpenHands—against the same axes and reports only partial coverage across all of them.

Pass-by-reference is the heavy-lifting capability: arguments arrive as live Python objects; the model sees only a bounded preview with the concrete type, actual length, and a head/tail sample. A list is rendered while the full variable remains in the REPL. Context is split into a static prefix that can be cached, an append-only event history, and dynamic tail blocks, which together preserve the KV-cache.