arXiv:2610.33588v1 Announce Type: new Abstract: As AI coding agents become more autonomous, developers increasingly direct, coordinate, and check their work. Yet little is known about these practices in everyday professional development. We shadowed 11 professional developers across approximately 15 hours of work, coded their activities and task transitions, analyzed 255 submitted prompts, and conducted 4.6 hours of follow-up interviews. Developers spent an average of 33.4% of observed time formulating prompts or checking AI responses, with task-switching rates ranging from 0 to 101.6 switches per hour. Beyond producing artifacts, prompts supported understanding, requirements clarification, agent configuration, and coordination. We identified three recurring patterns: \textit{Supervisor}, involving direct artifact inspection; \textit{Delegator}, involving dialogue-based coordination and delegated implementation and checking; and \textit{Learner}, involving agents primarily to build understanding. Developers sought technical information from agents and organizational context, authority, and alignment from colleagues. Our findings suggest future studies of how developers manage attention across concurrent agent tasks and tools that support verification, coordination, and learning.