NVIDIA Labs has open-sourced NOOA (NVIDIA Object-Oriented Agents), a model-agnostic Python framework designed to simplify how developers build and manage AI agents.
The framework takes a different approach from traditional agent development. Instead of separating prompts, tool definitions, callbacks, state management, and workflow graphs across multiple parts of a project, NOOA brings the core components together inside a single Python class.
In NOOA, methods represent actions an AI agent can perform, fields represent its state, docstrings function as prompts, and type annotations act as contracts enforced by the runtime. Developers can also distinguish between AI-driven and deterministic behavior: a method containing ... is completed dynamically through an LLM-driven loop, while a method containing normal Python code remains deterministic and can be called by the model as a tool.
NVIDIA reports strong results across several benchmarks, including 82.2% on SWE-bench Verified, 86.8% on CyberGym L1, and 85.1% mean RHAE on ARC-AGI-3. The company also reports that NOOA achieves these results with roughly half the token usage of the open harnesses used for comparison.
Is NVIDIA NOOA Ready for Deployment?
NOOA is available for developers, but NVIDIA makes its current limitations clear.
The framework is released under the Apache 2.0 license and can be installed using pip install nooa. Version 0.0.8 was released on July 30, 2026, and the project requires Python 3.12–3.13.
PyPI currently classifies NOOA as alpha software, while NVIDIA describes it as a research preview. This means organizations should carefully evaluate the framework before using it for critical production workloads.
One of the most important considerations is security. NOOA can execute code generated by an LLM. NVIDIA explains that its AST checks and module deny-lists should be considered defense-in-depth protections rather than a true security boundary.
For that reason, AI-generated code should run inside OS-level isolation, such as a container, virtual machine, or NVIDIA OpenShell.
Models are also interchangeable. NOOA uses LiteLLM, allowing developers to connect hosted APIs as well as local or self-hosted models through platforms such as Ollama and vLLM.
Potential users include AI-native startups, mid-market technology companies, enterprise AI teams, and applied-research organizations conducting agent evaluations or pilots. Industries that could benefit include developer tooling, cybersecurity, cloud and DevOps, analytics, financial-services operations, and customer support.
NOOA Treats an AI Agent Like a Python Object
The central idea behind NVIDIA Object-Oriented Agents is straightforward: an AI agent can be represented as a Python object.
Traditional agent systems often require developers to manage separate prompt templates, tool schemas, callback functions, state systems, and workflow definitions. NOOA combines these concepts within a Python class.
For example, methods define what the agent can do, fields store information about its current state, and docstrings provide instructions to the model. Type annotations provide additional structure by defining what inputs and outputs are expected.
The framework supports two primary execution strategies.
PredictStrategy performs a single typed LLM call and includes a local retry mechanism when the generated response fails validation.
CodeActStrategy, meanwhile, takes a more iterative approach. The model works through a Python REPL and can repeatedly execute code using execute_python(...). The process continues until the model calls return_result(...), after which the returned value is checked against the method’s return annotation.
This gives developers a way to combine traditional Python programming with model-driven reasoning inside the same interface.
Six Capabilities Unified in One Framework
The research team behind NOOA identifies six model-facing capabilities that it says the framework combines in one system:
- Typed inputs and outputs
- Passing data by reference
- Code as an action mechanism
- Programmable loop engineering
- Explicit object state
- APIs that the model itself can call
NVIDIA evaluated 14 frameworks and agent harnesses, including LangGraph, Google ADK, PydanticAI, smolagents, Claude Agent SDK, OpenAI Codex, and OpenHands.
One of NOOA’s key design choices is pass-by-reference data handling.
Instead of placing an entire large data structure into the model’s context, arguments remain as live Python objects. The model receives a limited preview containing information such as the object’s concrete type, actual length, and head/tail samples.
For example, a list containing 100 elements can be represented in approximately 30 tokens while the complete list remains available inside the REPL.
NOOA also divides its context into a cacheable static prefix, an append-only typed event history, and dynamic blocks at the end. This architecture is designed to preserve KV-cache reuse between model calls.
Built-In Memory Support
NOOA also includes an optional memory subsystem that can be attached to an existing agent without modifying the original agent itself.
The memory system provides seven model-callable tools for storing and retrieving records. These records are ranked using ACT-R activation, while the underlying information is stored in a single human-inspectable SQLite database.
This approach gives developers a relatively transparent way to inspect an agent’s stored information rather than relying entirely on an opaque external memory service.
NOOA Performance and Benchmark Results
NVIDIA’s evaluation covered 88 tests, each run five times across 10 models. In total, 4,309 of 4,400 records passed, producing a 97.9% pass rate.
A smaller stress-test group covering batching, error recovery, and decomposition achieved an 84.7% pass rate. The research found that the performance gap between smaller and frontier models becomes substantially larger under these demanding workloads, increasing from 3.2 points to 23 points.
The framework also produced notable end-to-end benchmark results.
On SWE-bench Verified, a 253-line NOOA agent achieved 82.2% with GPT-5.5 at xhigh effort, compared with 78.6% for OpenCode and 78.2% for PI. With Opus 4.6, the same framework reached 79.8%.
On Terminal-Bench 2.0, NOOA achieved 73.0% at high effort, compared with 60.7% and 68.5% for the other compared systems. PI performed better at xhigh effort, reaching 75.3%.
NOOA also achieved 86.8% on CyberGym L1 with network access disabled, which NVIDIA reports as the leading open-source result in its comparison.
For ARC-AGI-3, an agent using a one-page world-model skill achieved 50.2% mean RHAE with GPT-5.5 and 85.1% with GPT-5.6-sol, with the reported cost remaining below $20 per game.
Token Efficiency May Be NOOA’s Biggest Advantage
Beyond benchmark scores, NOOA’s efficiency is particularly notable.
For SWE-bench Verified, the framework reached 82.2% while using approximately 1.1 million tokens and 28 model calls per task. NVIDIA compares this with PI, which used approximately 2.2 million tokens and 66 calls while achieving 78.2%.
The difference is partly related to how agents terminate successfully.
According to the trace analysis, OpenCode can stop when a model responds without making a tool call. NOOA instead requires a typed TaskResult containing evidence and a verification command.
This gives the system a more explicit definition of what constitutes a completed task.
What NVIDIA NOOA Means for AI Agent Development
NOOA’s biggest contribution may be its attempt to make AI agents feel more like conventional software.
Instead of maintaining separate systems for prompts, state, tools, and orchestration, developers can express much of the agent’s behavior through familiar Python structures. This could make agents easier to test, trace, refactor, and version-control.
However, NOOA is still an alpha research preview, and its ability to execute LLM-generated code means security isolation remains essential.
For developers experimenting with autonomous coding agents, cybersecurity automation, infrastructure tools, analytics workflows, and multi-agent systems, NOOA provides an interesting alternative to graph-based agent frameworks.
The framework’s combination of object-oriented Python, typed contracts, live data references, model-driven code execution, explicit state, and relatively strong benchmark performance makes it a notable development in the rapidly evolving AI agent ecosystem.
Discover more from AiTechtonic - AI & Informative News
Subscribe to get the latest posts sent to your email.