When the Canvas Starts Acting, Who’s Really in Control?
1. The interface shifts from a manipulation screen to an active subject
- Agentic AI does not wait for the user's input; instead, it interprets context and suggests or executes the next action.
- If conventional UI was a surface that helped with clicks and input, agentic UI is closer to a participant that understands goals and continues the task.
- Designers' attention also shifts from the arrangement of screen elements to how far the system can judge and act.
2. A blank prompt alone makes it hard to know what is possible
- A blank input field may seem free, but it makes users explore functions and next actions on their own.
- Appropriate suggestions and feedback are needed so that possible tasks, the current state, and the scope of execution can be understood on screen.
- Nielsen's principles of visibility, user control and freedom, and error recovery also expand into agent control methods such as interruption, undo, and rollback.
3. The canvas becomes a space for exploring and connecting context
- In environments such as Miro, Figma, and Google Gemini, AI does not simply process the given data; it finds the necessary context and connects it to the task.
- Users can understand the flow and trust it only when the system reveals what it saw and why it acted.
- The movement of attention and the amount of information revealed also need to be designed together so that automatically continued tasks do not scatter the user's focus.
4. Control is designed through policy and handoff points
- Human-in-the-loop is closer to a structure that selects the moments when human judgment is needed, rather than approval at every stage.
- Human intervention can be reserved for major decisions or irreversible actions, while repetitive and recoverable processes can be left to the agent.
- The designer's role expands toward designing the points where humans and AI hand control back and forth, along with operating rules, constraint conditions, and levels of permission.
