Software as Content: Dynamic Applications as the Human-Agent Interaction Layer

Why Read This from an HCI Perspective

From an HCI perspective, this article raises the question of whether we need to redesign ‘the output of conversational AI’ not as a product feature, but as the interaction medium itself. In particular, it explains in concrete terms why chat becomes inefficient for tasks involving structured information, iterative refinement, and accumulated state. For UX practitioners, it offers criteria for deciding when an interface that users can manipulate is more appropriate than text when integrating AI capabilities into products. For researchers, it proposes new research agendas around dynamic interfaces and state management.

CIT's Commentary

From CIT’s perspective, SaC is an extension of GenUI while also marking an important divergence. If existing GenUI produces ‘well-presented responses,’ SaC extends that idea into a workspace where users can rework the response and create the next state. This can be seen as a re-interpretation in the AI era of principles long discussed in HCI—direct manipulation and cognitive externalization. However, in real products, not every task can be absorbed into a structured interface; there are clearly segments where natural language is more efficient for exploration, exception handling, and intent expansion. Accordingly, CIT interprets SaC not as a ‘replacement for chat,’ but as an adaptive interaction architecture that dynamically switches between text and structured manipulation depending on the task type. The key going forward is less about generatability than about how to design for state persistence, transition costs, and user control.

Questions to Consider While Reading

• Q.In SaC, what are the minimum requirements for an interface whose state persists—and at what point does the user begin to perceive the interface as an actual ‘tool’?
• Q.When structured manipulation and natural-language input coexist, how should the system propose switching criteria between the two channels to minimize users’ cognitive burden?
• Q.When evaluating the effectiveness of SaC-based systems, beyond task completion time, what metrics are appropriate for measuring accumulated state, personalization, and the sense of control?