A study with 19 developers produced a structured framework for how people supervise AI coding agents. The researchers identified six recurring activities: setting up tasks, tracking progress, reviewing outputs, stepping in to redirect, recovering from mistakes, and managing what the agent can touch. These are not abstract categories. They reflect specific moments where a developer decides, approves, or takes over, and where the design of the tool either helps or fails them. The framework reveals a gap between what giving work to an agent feels like and what it requires. Passing a task to an agent looks like a single action. In practice, a developer must first shape what the agent understands, then decide how much to watch while it runs, then judge what it returns. Each of those steps demands something different from the interface. A tool that handles the initial assignment but ignores supervision leaves the developer without what they need at the moments that matter most. For teams building agent products, this research names the pressure points. A developer does not stop caring about a task once they pass it on. They need to know the agent's progress without opening it, redirect it without restarting it, and judge its output fast. The study gives those needs a shared vocabulary, which makes it easier to spot where a product is thin.