NVIDIA NOOA Turns an AI Agent Into One Python Class

NVIDIA's Answer to Agent Framework Sprawl Is to Delete Most of It

Agent frameworks have accumulated a lot of machinery over the past two years: graph builders, node registries, planner abstractions, memory stores, orchestration DSLs. NVIDIA Labs has open-sourced a framework called NOOA that argues most of that is unnecessary, and it makes the argument by collapsing an entire agent into a single Python class.

• NOOA — NVIDIA Object-Oriented Agents — is a model-agnostic Python framework released as a research preview with code and benchmarks on GitHub
• Methods become the agent's available actions, fields become agent state, docstrings become prompts, and type annotations become contracts the runtime enforces
• A 253-line agent reaches 82.2% on SWE-bench Verified with GPT-5.5 at extra-high effort, and 86.8% on CyberGym L1 with network access blocked
• It uses roughly 1.1 million tokens per SWE-bench Verified task against about 2.2 million for competing harnesses

How the Object-Oriented Mapping Works

The core idea is that Python already has the vocabulary an agent needs, so the framework should reuse it rather than reinvent it. Write a method and it becomes a tool the model can call. Declare a field and it becomes state that persists across the agent's turns. Write a docstring and it becomes the prompt describing that capability. Annotate the types and the runtime enforces them as a contract, rejecting malformed calls before they reach your code.

The practical effect is that the agent definition and the agent implementation are the same artifact. There is no separate schema file describing tools that can drift from the functions those tools actually invoke, which is one of the more persistent sources of quiet breakage in agent codebases.

Is a Smaller Harness Actually Better?

The benchmark results suggest the reduction is not costing capability. A 253-line agent reaching 82.2% on SWE-bench Verified is competitive with far more elaborate scaffolding, and the 86.8% result on CyberGym L1 with network access disabled is reported as the top open-source figure on that evaluation.

The number that deserves the most attention, though, is token consumption. NOOA reports roughly 1.1 million tokens per SWE-bench Verified task where competing harnesses use around 2.2 million. Halving token spend while scoring higher is the kind of result that changes what is economically viable to run, and it points at a real cost of heavyweight frameworks: every abstraction layer that reformats context, re-summarises history, or re-injects instructions is paying tokens for the privilege.

Availability and Governance

NOOA is available now as a research preview on NVIDIA's GitHub, and NVIDIA is contributing it to the Open Secure AI Alliance, the industry coalition it convened around AI security and safety. Putting an agent harness under shared governance rather than keeping it purely in-house is a sensible move given how much of the current agent security surface lives in the harness rather than the model — a theme we explored when NVIDIA released a scanner for agent skills, covered in our piece on SkillSpector.

It also lands in a busy month for coding harnesses. Meta shipped its own multi-agent coding system in early August, which we wrote up in our Muse Code and Muse Spark 1.2 coverage. The contrast is instructive: one approach builds out observability and replay tooling around a rich harness, the other strips the harness down until there is less to observe.

Worth Trying If You Maintain Agent Code

For teams already running agents in production, the migration question is whether your existing tool definitions map cleanly onto plain methods. If they do — and for most tool-calling agents they will — NOOA is a low-risk experiment that could meaningfully cut both your line count and your inference bill. Being model-agnostic means you are not committing to a particular provider to find out.

More agent framework news in our AI section.

Sources: MarkTechPost — August 7, 2026; The New Stack — August 2026; The Hacker News — July 2026.