“The ARC-AGI V3 benchmark is designed to measure "agentic intelligence" by placing an AI in a novel, game-like environment where it must discover goals and controls through trial and error.”