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

Nvidia’s research argues that in long-horizon AI tasks, the surrounding harness matters more than the base model. By adding a custom memory-aware harness and a supervisor component, Nvidia says Claude Opus 5 reached a 100% score on ARC-AGI-3, versus 30% without the harness. The piece reinforces a broader market theme: agent performance, cost, and safety may depend as much on orchestration, tools, and runtime layers as on frontier model quality. It also highlights that open agent stacks could become a key competitive differentiator in enterprise AI infrastructure, potentially benefiting companies building tooling around models rather than only the model labs themselves.