Proactive agents should execute high-level user intent and allow inspection of work afterward, rather than following a task-by-task model or acting without user visibility, requiring a balance between autonomy and user control.