Advances in AI have transformed decision-support systems from simple tools for humans into collaborative teammates. As humans increasingly supervise multiple heterogeneous agents, effective collaboration depends not only on AI capabilities but also on human ability to understand them. While prior research has focused on improving AI accuracy, transparency, and explainability, comparatively little is known about how humans develop cognitive representations of multiple AI teammates and use them to supervise complex teams. This dissertation investigates how humans develop and apply mental models of AI, i.e., Team Models, in hierarchical Human-Agent Teams (HATs). Through human-subject studies for human-agent hierarchical triads, this dissertation examines whether Team Models improve supervisory decision making, the factors that influence their development, and how different forms of prior information about AI shape Team Model formation and joint outcomes. The findings demonstrate that Team Models improve supervisory coordination, reduce workload, and enhance task efficiency, particularly when AI teammates provide conflicting recommendations. However, Team Model development is constrained by inter-agent dependencies, users' limited ability to identify AI failure modes, and the abstraction level of information presented by the system. The research further shows that prior information about teammate characteristics only selectively improves Team Model development, with its effectiveness depending on team structure. Moreover, increased knowledge about AI teammates does not necessarily produce calibrated reliance or improved collaboration, as users often rely on environmental feedback and simple heuristics rather than reasoning about teammate capabilities. Overall, this dissertation extends shared mental model theory hierarchical HATs and provides human-centered design guidelines that support better supervision and collaborative decision making.