Natural and artificial societies often divide the workload between specialized members. For example, an ant worker may preferentially perform one of many tasks such as brood rearing, foraging and nest maintenance. A robot from a rescue team may specialize in search, obstacle removal, or transportation. Such division of labor is considered crucial for efficient operation of multi-agent systems and has been studied from two perspectives. First, scientists address the "how" question seeking for mechanical explanations of division of labor. The focus has been put on behavioral and environmental factors and on task allocation algorithms leading to specialization. Second, scientists address the "why" question uncovering the origins of division of labor. The focus has been put on evolutionary pressures and optimization procedures giving rise to specialization. Studies have usually addressed one of these two questions in isolation, but for a full understanding of division of labor the explanation of the origins of specific mechanisms is necessary. Here, we rise to this challenge and study three major transitions related to division of labor. By means of theoretical analyses and evolutionary simulations, we construct a pathway from the occurrence of cooperation, through fixed castes, up to dynamic task allocation. First, we study conditions favoring the evolution of cooperation, as it opens the doors for the potentially following specialization. We demonstrate that these conditions are sensitive to the mechanisms of intra-specific selection (or "selection methods"). Next, we take an engineering perspective and we study division of labor at the genetic level in teams of artificial agents. We devise efficient algorithms to evolve fixed assignments of agents to castes (or "team compositions"). To this end, we propose a novel technique that exchanges agents between teams, which greatly eases the search for the optimal composition. Finally, we take a biological perspective and we study division of labor at the behavioral level in simulated ant colonies. We quantify the efficiency of task allocation algorithms, which have been used to explain specialization in social insects. We show that these algorithms fail to induce precise reallocation of the workforce in response to changes in the environment. We overcome this issue by modeling task allocation with artificial neural networks, which lead to near optimal colony performance. Overall, this work contributes both to biology and to engineering. We shed light on the evolution of cooperation and division of labor in social insects, and we show how to efficiently optimize teams of artificial agents. We resolve the encountered methodological issues and demonstrate the power of evolutionary simulations to address biological questions and to tackle engineering problems.