This dissertation is concerned with the modeling and analysis of large-scale online platforms, with a focus on the complexities that arise from their multi-agent nature. Specifically, this thesis is interested in two aspects of platform operations: (i) understanding how current models fail to take into account important facets of human behavior and interactions, and leveraging these insights to improve upon state-of-the-art algorithms; and (ii) designing algorithms toward the social good. In broaching the first topic, this thesis illustrates the perils of the strong informational and rationality assumptions typically imposed on human behavior in a variety of settings, and, using tools from online learning and stochastic control, proposes simple and intuitive solutions with strong performance guarantees. In regards to the second area of focus, this work considers how ride-hailing services can be leveraged in conjunction with more traditional transportation options for social welfare maximization objectives, and tackles operational and market design aspects of this problem. In each of these areas of focus, we develop efficient algorithms with provable guarantees that outperform state-of-the-art methods on real-world datasets.