As AI agents move beyond responding to prompts and begin actively navigating context, modifying files, running workflows, and making decisions, interface design is shifting from specifying every user action to orchestrating how responsibility moves between people and systems. The key design challenge becomes deciding what an agent can do autonomously, when it should ask for confirmation, what must remain reversible, and which judgment-heavy decisions should always return to a human. In this model, designers increasingly shape the product's policy layer—behavior, permissions, constraints, defaults, and attention—while human value moves toward deciding what deserves oversight and what can safely be delegated.