The Three Stages of Agent UI
In an era when AI writes frontend code better than people, why are UIs still static? That's the question Ruben Casas of Postman posed at the AI Engineer conference. He frames the evolution of agent UI in three stages: static, where agents fill data into developer-prebuilt components; declarative, where agents write JSON descriptions that a rendering engine maps onto the design system; and generative components, where the model writes HTML and CSS on the spot at runtime.
Why the Current Equilibrium Is Declarative
The declarative approach preserves design-system consistency while enabling personalization, and it's cheap. What blocks generative isn't capability — it's trust. Code an LLM writes on the fly can't be shipped to a user's screen unverified, so a safe distribution structure — sandboxing — has to come first.
Why MCP Apps Are Getting Attention
MCP apps come with authentication, tool invocation, and a double-iframe sandbox by default, so they can serve as the distribution channel for generative UI. Some read Anthropic implementing its own visualization features as MCP apps — rather than native rendering — as a strategic signal.
The TV Analogy — the Limits of Bolting On a Chatbot
When TV first arrived, people just pointed cameras at radio shows. SaaS products bolting chatbot windows onto their homepages today are at exactly that stage. The next stage is collaborative UI, where humans and agents edit the same canvas together.
The same question applies when designing enterprise AX: not where to attach a chatbot, but where to put the interface where agents and people work together. This is why SH Consulting treats the interface layer as its own concern in AX design.