Individual neurons are interconnected, resulting in coordinated patterns of activity and emergent population dynamics that underlie complex behavior. I investigated the time-varying co-activity of neural populations, summarized as functional networks (FN), and their relation to motor behavior. First, I found that the structure of FNs in the primary motor cortex of macaques performing an instructed-delay reaching task was specific to the instructed reach, and that FNs constructed from trials with closer reach directions are also closer in network space. I extended this analysis by computing co-activity within a short interval across time to construct temporal FNs. This revealed that reach-specific differences in FNs emerge shortly after instruction and, consequently, become decodable for reach direction. In fact, FNs provide an additional source of information about behavior beyond what is carried by firing rates alone. Next, in the sensorimotor cortex of a marmoset performing a self-initiated virtual prey capture task, FNs at specific task events (such as trial start, target presentation, etc.) can be discriminated based on its nearest neighbors using either a low-dimensional metric (distance in an embedding space) or a network alignment score. Partitioning temporal FNs based on structure showed FN states that corresponded to specific behaviors such as arm extension or licking, and that successful trials involve a stereotyped sequence of these states. These results suggest that FNs can provide novel insight on the statistical regularities of neural co-activity that produce neural population dynamics during voluntary goal-directed reaching behavior