Nvidia: The Agent Harness, Not the Model, Is the Real Hero

Nvidia published new research on Friday that flips a common assumption about AI agents: the model you pick matters far less than the scaffolding around it. In experiments on long-horizon tasks, a carefully engineered "harness" took Claude Opus 5 from a 30% score to a perfect 100% on the ARC-AGI-3 interactive reasoning benchmark, according to TechCrunch.

What Nvidia found

Nvidia's researchers built a custom harness called Agentic Variation Operators (AVO), tuned for memory management and equipped with a "supervisor" component that nudges the working agent back on track when it drifts. With that harness, Opus 5 hit 100% on ARC-AGI-3. Without it, the same model scored 30%, which was still the top result among every model tested.

The benchmark, built around 2D games with no instructions, has particularly irked OpenAI. The company's models scored below 10%, prompting OpenAI to run its own research last month. Like Nvidia, OpenAI found that tweaking just two harness settings tripled its scores, but no other lab came close to a perfect result.

The harness is the agent

"Generally speaking the world interprets an agent almost as an API of the model," Adel El Hallack, vice president of product in Nvidia's AI unit, told TechCrunch. "It is the model. It is the scaffolding around the model, which we call the harness. It is the runtime and the associated skills and libraries that we give it access to."

The supervisor layer was the breakthrough. "The more interesting part was introducing a supervising agent in addition to your main agent that's doing the work," El Hallack said. "It almost acts like a CEO to nudge the agent when it goes off direction."

Nvidia is not shipping AVO as a product. Instead, the company produces open building blocks for building harnesses under its NeMo brand, making the larger point that open harnesses give builders more control over accuracy and security than closed stacks.

What it means for builders

For teams building and deploying agents, the takeaway is concrete: stop obsessing over which frontier model to use and start investing in the harness. Memory, context management, feedback loops, and a supervisor layer are what turn a model into an agent capable of long-horizon work.

That shift is already visible across the industry, as agents move into the workplace and the focus turns from raw model capability to orchestration, tooling, and runtime control.