Nvidia’s NOOA makes an agent one Python class

Nvidia dropped a new framework last week called NOOA, short for Object-Oriented Agents, and it's making a simple but disruptive bet: the way you structure an agent around a model matters just as much as the model itself. Instead of spreading prompts, tool schemas, callbacks, and workflow logic across five different files and three different libraries, NOOA squeezes all of it into a single Python class. Docstrings become prompts. Type annotations become enforced contracts. Methods with a body of just three dots get filled in by an LLM at runtime, while normal methods run as plain, boring, deterministic code.

That's a real answer to a mess a lot of teams know well. Adnan Masood, chief AI architect at UST, told The New Stack that today's typical setup — Jinja templates for prompts, JSON schemas for tools, Python for callbacks, and a graph abstraction for the workflow — is fragmented by design. NOOA collapses that sprawl into something a human or an AI coding assistant can actually read top to bottom, using tools developers already know. Nvidia is contributing the project to the Open Secure AI Alliance, the industry group it launched the same week to push open-source agent tooling.

The numbers Nvidia is showing off are not small. On SWE-bench Verified, NOOA paired with GPT-5.5 hits 82.2% using 29 LLM calls and about 1.1 million tokens per task. Comparable harnesses needed either 66 calls and 2.2 million tokens to reach 78.2%, or the same 29 calls but 1.3 million tokens for 78.6%. Nvidia calls that parity or better at roughly half the cost. It also reports 86.8% on the CyberGym L1 vulnerability benchmark and 50.2% mean RHAE on ARC-AGI-3.

But the same feature that makes NOOA inspectable — one class, one trace, one readable structure — also concentrates the blast radius when things go wrong. Masood points out that NOOA lets the model act by writing and running its own Python, which is powerful but also raises the stakes of a prompt injection buried in some webpage or document. Karthik Karunanithi at IBM raises a subtler concern: because a runtime-filled method and a hardcoded one share identical signatures and docstrings, telling deterministic code apart from probabilistic behavior during a review gets genuinely harder. Siddhartha Saxena of Thine and Merlin AI puts the readability tradeoff more bluntly, noting that anything readable to humans is readable to attackers too. Nvidia's answer is to pair NOOA with its OpenShell secure runtime for production use, plus sandboxing and scoped credentials, but that's a mitigation, not a solved problem.

What's clear is that Nvidia isn't just chasing leaderboard bragging rights. Distinguished research scientist Paul Furgale framed the release as part of a bigger push: open AI, he argued, needs open research into how models actually interact with computers, not just open weights. Whether the developer community adopts NOOA wholesale or just borrows its ideas, the framing itself — that harness design alone can swing benchmark results by double digits — is likely to stick around longer than any specific class definition.