Modularity is a fundamental organizing principle in both biological and artificial intelligence systems. In the brain, functionally specialized neural circuits divide into dorsal (spatial/action-oriented processing) and ventral (object recognition) pathways. In AI, Mixture-of-Experts architectures selectively activate specialized subnetworks. Despite these parallels, we lack a fundamental understanding of why such modularity emerges and how to design effective modular systems. This project systematically investigates the emergence and functional significance of modular organization across biological and artificial vision systems, using a comparative approach spanning multiple species (mouse, monkey, human) and model architectures. We test three hypotheses: the Task-Tension Hypothesis (computational tensions naturally drive modularity in self-supervised models), the Perspective-Induced Hypothesis (egocentric vs. allocentric processing drives specialization), and the Agent-Action Hypothesis (action requirements drive dorsal/ventral-like division in RL agents).