Most AI agents still live in a chat window, terminal, messaging app, or fixed web interface. They can answer questions, edit files, and run commands, but the interface stays the same.
I think the agent should create the view needed for the current task.
A summary can stay as text. A data comparison should become a table. Monitoring needs a dashboard, and project work may need a small task board.
Those views can be part of the interaction instead of separate applications built in advance.
We are used to this model:
fixed app UI + chatbot bolted on topAnother model is:
agent + dynamic workspace + generated widgets + persistent state + rollbackThe agent shapes the working environment around the task instead of forcing every result through chat.
Software that changes its own interface can break it. Undo, audit history, permissions, sandboxing, and recovery need to be part of the design.
Chat still works for commands and discussion. The generated workspace should hold the state and controls that would otherwise be squeezed into prose.