Multi-agent activity is an emergent process, with the roles and responsibilities of individual actors a self-organized consequence of task-dynamic constraints and perturbations. The shepherding paradigm, first investigated by Nalepka, Kallen, Chemero, Saltzman, and Richardson, (2017), was directed towards exploring the emergence of stable multi-agent behavioral modes within dynamically changing task-environments. The task involved pairs of participants using their hands to contain a herd of autonomous and reactive “sheep” within a virtual game field projected on a tabletop display. Initially, all participants employed a search-and-recover (S&R) mode of behavior, moving from sheep-to-sheep to corral the herd to a target containment region in the center of the game field. However, a subset of dyads learned to coordinate their movements, forming an oscillating “wall” that contained the herd (termed coupled oscillatory containment)—a behavioral mode termed coupled oscillatory containment (COC)—which allowed dyads to achieve superior task performance. Experiment 1 investigated a potential control parameter to promote the emergence of COC behavior, as well as determine whether changes in oculomotor behavior might predict its emergence using recurrence quantification analysis (RQA). Experiment 2 sought to validate weather S&R and COC behavior also defined a more realistic herding situation that involved the full-body movement of participants in a large task space. Results indicated that manipulating task difficulty, by controlling how fast the sheep could move, promoted the use of more coordinated modes of behavior (Experiment 1 and 2). Significant changes in the determinism and complexity (entropy) of oculomotor behavior was observed two trials prior to the discovery of COC behavior using RQA (Experiment 1). In full-body herding (Experiment 2), a subset of dyads discovered a coordinated mode of behavior that involved a joint circling pattern in a fixed direction, with the frequency with which this behavioral mode emerged increasing with increases in task difficulty. Future work will investigate the informational basis that leads to the discovery of optimal solutions (i.e., oscillatory behavioral coordination) in the shepherding task, as well as the design of flexible, adaptive artificial (machine) systems that can learn and work alongside humans in these unstable task-environments.