NOOA shifts attention from model size to harness design

NVIDIA’s new NOOA work is a direct argument that agent performance depends as much on the surrounding software architecture as on the model selected for a run. The headline numbers are unusually concrete: NVIDIA’s technical blog says NOOA reached 82.2% on SWE-bench Verified with GPT-5.5, 86.8% on CyberGym L1, and 85.1% mean RHAE on ARC-AGI-3 with GPT-5.6-sol.

“We're contributing open models, weights, data and research to the Open Secure AI Alliance. This includes NVIDIA Labs Object-Oriented Agents (NOOA): open agent harness research in the form of an object-oriented framework that reaches state-of-the-art results across software engineering, cybersecurity and reasoning benchmarks.”

The source tweet from NVIDIA AI frames NOOA as a contribution to the Open Secure AI Alliance. @NVIDIAAI normally publishes developer-facing model, infrastructure, and benchmark updates, so the tweet is best read as a technical release pointer rather than a broad consumer message.

NOOA’s central design move is to treat an agent as a single Python class. Methods become capabilities, fields become state, docstrings become prompts, and type annotations become enforced contracts. A method body written as an ellipsis is completed at runtime by an LLM-driven loop, while ordinary Python methods remain deterministic code. That makes the agent closer to conventional software: reviewable, testable, traceable, and versioned.

The efficiency claim is as important as the benchmark score. NVIDIA says NOOA reached 82.2% on SWE-bench Verified using 29 LLM calls and roughly 1.1 million tokens per task. Comparator harnesses needed either 66 calls and 2.2 million tokens to reach 78.2%, or 29 calls and 1.3 million tokens to reach 78.6%. If external users reproduce that pattern, the practical lesson is that context representation and state passing can change both accuracy and inference cost.

The blog also describes a long-term memory subsystem stored in a human-readable SQLite file, with typed records, importance values, tags, and relationships such as supports or contradicts. That detail matters because many deployed agents still rely on ad hoc summaries or prompt stuffing. The next thing to watch is whether the public NOOA code and evaluation scripts let independent teams reproduce the SWE-bench, CyberGym, and ARC-AGI-3 results under the same constraints.

#nooa

NVIDIA Opens NOOA Agent Harness With 82.2% SWE-bench Score