What Is NOOA?
<cite index="2-1">NVIDIA Labs has open-sourced NOOA (NVIDIA Object-Oriented Agents), a model-agnostic Python framework for building Artificial Intelligence (AI) agents.</cite> <cite index="18-1">The project was released on GitHub under an Apache 2.0 licence.</cite> <cite index="1-4">Announced on July 27, 2026, NOOA offers an open-source, Python-based architecture designed to enhance agent efficiency, accuracy, and cost-effectiveness across industries including software engineering, cybersecurity, and general reasoning.</cite>
Core Design
<cite index="3-6">An agent harness is the software layer around a model that renders context, executes actions, manages state, and decides when a task is done.</cite> <cite index="2-2,2-3">Agent development today is split across prompt templates, tool schemas, callback code, and workflow graphs — NOOA collapses all of it into one Python class.</cite> Under this object-oriented model, <cite index="3-9,3-10,3-11">fields store an agent's state, methods expose its capabilities, docstrings act as prompts, and type annotations define the contracts the model must follow. A method containing an ellipsis body is completed at runtime by a Large Language Model (LLM)-driven loop, while a method containing ordinary Python remains deterministic code.</cite>
<cite index="3-12">The same structure lets developers use familiar testing, tracing, version control, and refactoring workflows instead of splitting agent behavior across prompts, tool schemas, callbacks, and workflow graphs.</cite> <cite index="9-4">Context, events, state rendering, long-term memory, and validated LLM loops are presented through Python-style APIs rather than a separate agent-specific language.</cite>
Benchmark Performance
<cite index="1-7">NVIDIA claims that NOOA can deliver double-digit percentage improvements in benchmarks while reducing token usage and operational costs by up to 50%.</cite> Specific evaluations support those claims: <cite index="2-12">a benchmark-agnostic 253-line agent reaches 82.2% on SWE-bench Verified with GPT-5.5 at extra-high effort, against 78.6% for OpenCode and 78.2% for PI, and 79.8% with Opus 4.6.</cite> <cite index="2-14">On CyberGym L1 it solves 86.8% with network access blocked — the top open-source result reported.</cite> <cite index="2-15">On ARC-AGI-3, one agent with a one-page world-model skill reaches 50.2% mean RHAE with GPT-5.5 and 85.1% with GPT-5.6-sol, under $20 per game.</cite>
Open Secure AI Alliance Context
<cite index="5-8">NVIDIA is contributing NOOA to the Open Secure AI Alliance, an industry group NVIDIA formed to build and share open-source and open-weight tools for AI development.</cite> <cite index="12-11">According to NVIDIA's announcement, the alliance brings together NVIDIA, Microsoft, IBM, Red Hat, Hugging Face, Cloudflare, CrowdStrike, Palo Alto Networks, Palantir, Databricks, Snowflake, Salesforce, ServiceNow, the Linux Foundation, and others.</cite> <cite index="11-8">NVIDIA tied the alliance's case for locally controlled defensive models to the July intrusion at Hugging Face, where an autonomous agent system compromised parts of the company's production infrastructure.</cite>
Notably, <cite index="14-9,14-10">OpenAI, Google, and Meta appear among the associated policy letter's signatories but are absent from the alliance's inaugural membership list, and Anthropic appears on neither list as of July 27, 2026.</cite> <cite index="13-4">The alliance has output but no published governance — member tooling shipped on day one, but analysts found no charter, no named governing board, no defined technical workstreams, and no delivery schedule.</cite>
Security Caveats
<cite index="7-3,7-4">LLM-generated code may take dangerous or unwanted actions, including sending private data to uncontrolled locations, deleting files, or modifying its environment. NVIDIA recommends running NOOA agents in a sandboxed environment isolated from the primary filesystem, such as NVIDIA OpenShell.</cite> <cite index="7-5,7-6">NOOA validates generated code through abstract syntax tree (AST) checks and applies module deny-lists before execution, though these are described as defense-in-depth guardrails, not a containment boundary.</cite> <cite index="11-4">A July 27 review of the public repository found a v0.0.6 tag dated July 22.</cite>