Nvidia research shows the wrapper around AI models can drive double-digit benchmark gains

Nvidia just published research that should make every AI company rethink where they’re spending their engineering hours. The finding: the “harness” wrapping an AI model, meaning the architecture that manages context, memory, and actions, can matter more than the model itself.

The company’s technical blog post, titled “Six Agent Harness Capabilities for Higher Model Performance,” lays out how harness design alone can produce double-digit percentage improvements in benchmark scores. At the same time, it dramatically reduces the number of tokens consumed. Same model, better packaging, radically different results.

What Nvidia actually built

Alongside the research, Nvidia Labs released an open-source framework called NOOA, short for NVIDIA Labs Object-Oriented Agents. Written in Python, the framework treats AI agents as individual classes, borrowing principles from traditional software engineering rather than the brute-force scaling approach that has dominated AI development.

NOOA incorporates typed input/output, pass-by-reference memory management, code-based actions, and model-callable APIs.

The benchmark results back up the approach. NOOA scored 82.2% on SWE-bench Verified, a widely used standard for evaluating AI coding agents, surpassing previous state-of-the-art marks. On CyberGym L1 tests, it hit 86.8% using general-purpose agents paired with models like GPT-5.5.

Perhaps most striking was the ARC-AGI-3 benchmark performance. NOOA achieved a 50.2% mean score with GPT-5.5 and jumped to 85.1% when integrated with GPT-5.6-sol. The cost per game with GPT-5.6-sol came in under $20.

Why the wrapper beats the engine

NOOA’s pass-by-reference memory management and elimination of context-compaction processes effectively halved token usage compared to previous designs.

Cutting token consumption in half is not just an engineering curiosity. Tokens are the unit of cost in AI inference. Every API call, every cloud compute bill, every enterprise deployment budget is denominated in tokens. Halving token usage while improving output quality is the kind of efficiency gain that changes business models.

What this means for the AI landscape

Nvidia releasing NOOA as open source is a deliberate strategic choice. The company sells the GPUs that power AI training and inference. Anything that makes AI cheaper and more accessible to deploy increases demand for Nvidia hardware across a broader customer base, not just among the handful of hyperscalers training frontier models.

The research effectively hands the open-source community a new set of tools to reproduce and extend Nvidia’s results.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.