The rule. Before a multi step agent acts, show the steps it intends to take, checkpoint the consequential ones, and keep a way to stop within reach.
Why. Two of Amershi et al.'s guidelines for human-AI interaction depend on the user seeing the work in time to react: make it easy to dismiss undesired AI services (G8), and make it easy to edit, refine, or recover when the system is wrong (G9) (Amershi et al., 2019). A plan shown before the run is what makes both possible mid run instead of after the damage. Visible staged progress also buys patience: Nielsen Norman Group cites a University of Nebraska-Lincoln study in which participants shown a moving progress indicator were willing to wait on average three times longer than those shown none, recommends showing the number of steps when duration is hard to estimate, and warns that without an option to stop, the design "may be hijacking control of the system, leaving the user powerless" (Sherwin, 2014).
Seen in the wild. Claude Code's plan mode has the agent research and propose its changes first; edits stay blocked until the user approves the plan (Anthropic, n.d.).