Nvidia published research on Friday arguing that the harness, not the underlying model, is what determines whether an AI agent can actually finish long-horizon work.
The demonstration is difficult to argue with. ARC-AGI-3 is a set of 2D games with no instructions. A model has to figure out how to play and win on its own, the way a person would sit down to an unfamiliar game with no manual, and the score is the percentage of games it actually wins. Running on its own, Claude Opus 5 won thirty percent of them, which was still the best result of any model tested that way. Running inside a custom harness that managed memory carefully and added a supervising agent to nudge it away from dead ends, that same model won one hundred percent. Same model, same test.
Nvidia’s vice president of product framed the real takeaway as control, explaining that open harnesses “allow you to turn a lot more knobs,” across the harness, the infrastructure, and the runtime.
If the harness is where cost and performance actually live, then the harness is what creates lasting, compounding value, and the model underneath it is what gets consumed and replaced the moment there’s a better one. Rippling governed how its requests were routed and cut its compute bill by sixty-three percent while processing 605 billion tokens a month, with no measurable loss of quality. That is not a technical optimization buried in an infrastructure team. That is where the compounding value in enterprise AI is won or lost. Satya Nadella compressed the whole strategy into five words: rent the model, own the harness.
Keep in mind that more knobs only becomes an advantage when someone in your organization owns and knows how to turn them.
If you are making platform decisions this quarter, this argument is Chapter 10 of my new book. It covers the architecture that makes “own the harness” operational, the difference between coupling you chose and coupling you inherited, and how to put a dollar figure on an exit cost before you sign the contract rather than after. Balancing the AI Value Equation: https://www.amazon.com/dp/B0HFHQRDM5