“In agentic AI systems performing long-horizon tasks, small initial performance differences between models are expected to compound, leading to significant disparities in final outcomes.”