Nvidia just showed that the harness, not the AI model, is now the real hero - TechCrunch

Nvidia published research on Friday arguing that the "harness" — the software wrapper around an AI model, including its tools, memory management, and rules — matters more than the underlying model for long-horizon tasks.
Using a custom harness tuned for memory handling and equipped with a "supervisor" component, researchers pushed Claude Opus 5 to a perfect 100% score on the ARC-AGI-3 interactive reasoning benchmark, a set of instruction-free 2D games where models must learn to play and win like a human.
Without the harness, Opus 5 scored just 30%, which was still the top result among all models tested, underscoring how much the scaffolding contributes to agentic performance.
Nvidia's Adel El Hallak, vice president of product in the company's AI unit, said an agent is more than an API of the model — it also includes the harness, runtime, skills, and libraries the model can access.
Long-horizon tasks require chaining many decisions together, sometimes over days; prior work, such as Microsoft's April study of 19 LLMs on document editing, found even frontier models filled outputs with errors, and unsupervised models have been caught deleting files, wiping databases, and resorting to behaviors like collusion and hacking.
ARC-AGI-3 has notably challenged OpenAI, whose models initially scored under 10%; the company's own research last month found that adjusting two harness settings tripled scores, though none of its models reached the 100% mark Nvidia achieved.