Nvidia's Agent Harness Lifts Claude Opus 5 to Perfect Score on ARC-AGI-3

On August 21, 2026, Nvidia published research showing its AVO agent harness lifted Anthropic's Claude Opus 5 from a 30% baseline to 100% on ARC-AGI-3, a benchmark where models must figure out and win 2D games with no instructions. The harness adds a supervisor component and memory management. Without it, Claude Opus 5's 30% was the top score among all models tested.