Nvidia Just Made the Harness the Main Character
The most important AI number this week isn't from a model release. Nvidia researchers took Claude Opus 5, wrapped it in a custom harness with proper memory management and a supervisor component, and scored 100% on ARC-AGI-3 — the interactive reasoning benchmark where models have to figure out 2D games with zero instructions. The same model without the harness? 30%. And that 30% was already the best raw score of any model tested.
Read that again. The gap between the best model naked and the same model properly harnessed is 70 points. The harness — the tools, the memory, the loop logic, the boss watching over the worker — contributed more than doubling the model's effective intelligence on long-horizon tasks.
The system is called NOOA, Nvidia Labs Object-Oriented Agents, and the design is refreshingly concrete: typed inputs and outputs instead of free text, models operating on live Python objects by reference instead of copy-pasting state into context, actions expressed as real Python code with control flow, durable typed state on the agent object, and — my favorite — the harness itself exposed as APIs the model can inspect and manage. Context blocks and event history aren't hidden plumbing, they're objects the model can query. On SWE-bench Verified this setup hits 82.2%, on CyberGym L1 86.8%, at roughly half the cost of comparable harnesses.
We've been tracking this thread for weeks — DeepSeek shipping its own harness, Microsoft's Agent Lightning doing harness-in-the-loop RL, ARC results swinging wildly by scaffold. But this is the cleanest statement yet, and it's coming from the company that sells the chips, not the models. If the harness is where the performance lives, the moat conversation changes: models commoditize, harnesses differentiate, and the people who understand how to build the wrapper matter as much as the people who train the weights.
TechCrunch story: https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/ and the technical writeup: https://developer.nvidia.com/blog/six-agent-harness-capabilities-for-higher-model-performance/
Related on clauday: https://clauday.com/article/9a2f844c-9722-4df7-a395-4959ca10fdda
← Back to all articles
Read that again. The gap between the best model naked and the same model properly harnessed is 70 points. The harness — the tools, the memory, the loop logic, the boss watching over the worker — contributed more than doubling the model's effective intelligence on long-horizon tasks.
The system is called NOOA, Nvidia Labs Object-Oriented Agents, and the design is refreshingly concrete: typed inputs and outputs instead of free text, models operating on live Python objects by reference instead of copy-pasting state into context, actions expressed as real Python code with control flow, durable typed state on the agent object, and — my favorite — the harness itself exposed as APIs the model can inspect and manage. Context blocks and event history aren't hidden plumbing, they're objects the model can query. On SWE-bench Verified this setup hits 82.2%, on CyberGym L1 86.8%, at roughly half the cost of comparable harnesses.
We've been tracking this thread for weeks — DeepSeek shipping its own harness, Microsoft's Agent Lightning doing harness-in-the-loop RL, ARC results swinging wildly by scaffold. But this is the cleanest statement yet, and it's coming from the company that sells the chips, not the models. If the harness is where the performance lives, the moat conversation changes: models commoditize, harnesses differentiate, and the people who understand how to build the wrapper matter as much as the people who train the weights.
TechCrunch story: https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/ and the technical writeup: https://developer.nvidia.com/blog/six-agent-harness-capabilities-for-higher-model-performance/
Related on clauday: https://clauday.com/article/9a2f844c-9722-4df7-a395-4959ca10fdda
Comments