Nvidia's newly released NOOA framework achieved the same performance as competing systems while using roughly half the tokens, according to benchmark results the company published last week. The Object-Oriented Agents framework, which Nvidia contributed to the Open Secure AI Alliance it formed last week, is built on a single principle: an AI agent is one Python class. The approach aims to address fragmentation in agent development by bundling capabilities, state, and prompts into a unified structure rather than scattering them across multiple files and abstractions.
On the SWE-bench Verified test, NOOA reached 82.2% accuracy with GPT-5.5 while making 29 LLM calls and consuming roughly 1.1 million tokens per task. Comparison systems required 66 calls and 2.2 million tokens to hit 78.2%, or 29 calls at 1.3 million tokens to reach 78.6%. On the CyberGym L1 vulnerability-rediscovery benchmark, NOOA solved 86.8% of tasks with GPT-5.5. The framework achieved 50.2% mean RHAE on ARC-AGI-3, a general reasoning benchmark.
Nvidia describes the framework's design in its announcement: "Its methods are its capabilities. Its fields are its state. Its docstrings are its prompts." A standard Python method whose body contains only an ellipsis gets completed at runtime by a language model loop, while methods with normal bodies run as ordinary, deterministic Python. The company argues that "harness design alone can account for double-digit swings in benchmark results and significant differences in token cost, with the same underlying model." Nvidia pairs NOOA with its OpenShell secure runtime for production deployment.
Industry experts who reviewed the framework see both promise and tradeoffs. Adnan Masood, chief AI architect at UST, told The New Stack that current agent development scatters components by design, with prompts in Jinja templates, tool definitions in JSON schemas, callbacks in Python, and workflow drawn as a graph in yet another abstraction. Consolidating those elements into one Python class makes agent code easier to audit than "logic smeared across prompt files and scattered scripts," according to both Masood and Siddhartha Saxena, co-founder of Thine and Merlin AI. But Saxena noted that if something's readable for humans, it's readable "for aliens, as well." Karthik Karunanithi, solution architect at IBM, questioned whether the framework might make it harder to distinguish deterministic code from probabilistic behavior, since both appear with the same signature and indentation. Masood highlighted another concern: the framework's use of code as action widens the blast radius of a prompt injection attack, where malicious text in a document or webpage steers the model.
Nvidia says it released NOOA because the future of open AI depends on more than open models — it also depends on open research into how those models interact with computers. Paul Furgale, distinguished research scientist at Nvidia, posted that while the company doesn't expect everyone to adopt NOOA, it's encouraging the developer community to adopt, challenge, and improve its techniques. The framework represents a bet that the harness around a model may matter as much as the model itself.