Harness beats model in Nvidia's AI research

Nvidia research shows the software harness around an AI model matters more than the model itself for long-horizon tasks. Using a custom harness with a supervisor component, researchers pushed Claude Opus 5 from a 30% to a perfect 100% score on the ARC-AGI-3 benchmark. The findings echo similar research from Databricks, which found harness choice can double AI costs regardless of model selection. Nvidia argues open harnesses give users far more control over accuracy and performance than most realize, and its researchers built a custom system called Agentic Variation Operators to prove the point.