Traditional frontend has become a bottleneck. For decades, software design assumed an inevitable adoption cost: forcing users to understand, memorize, and operate static graphical interfaces, full of menus, dashboards, and rigid navigation flows. Today, that cognitive toll is no longer necessary.

The popularization of applications like WhatsApp and ChatGPT standardized natural language-based interaction for users of all ages. We are all trained to express our intention directly, without visual intermediaries. However, proposing the “Zero UI” model exclusively as a plain text chat is a limiting approach. The true evolution is Generative UI: interfaces that don’t exist until they’re needed, assembled on the fly in response to user intent.

We are moving from “click interfaces” to “intent interfaces”, where we interact directly with business logic through an orchestrator that understands what we want and delegates execution to specialists.

The Mix of Experts Analogy

Within the most advanced language models exists an architecture called Mix of Experts (MoE): instead of activating the entire model for every query, an intelligent router analyzes the input and selects which specialized submodels should intervene. Not all experts participate in every response; only those relevant to that specific intention.

Zero UI applies exactly that same principle, but at system scale. The model acts as a router: it receives an intention expressed in natural language and decides which specialized tools should be executed, in what order, and with what parameters. Each tool is an “expert” with a concrete function — checking stock, issuing an order, generating a report — and the model orchestrates a precise choreography without the user even knowing these specialists exist.

The result is an interface that doesn’t impose a predefined structure on the user, but adapts to what they need at that moment.

Under the hood: Function Calling and Backend Reuse

For this paradigm to be viable at scale, we don’t build logic from scratch or create parallel systems. The technical key to Zero UI lies in reusing existing services in the backend — REST APIs, microservices, databases — by encapsulating them through Function Calling or Tools.

In this scheme, the model acts as an intelligent orchestrator. Let’s consider a concrete case: a user instructs the system “I need to restock the 5 best-selling products from last month”. Instead of forcing the user to navigate through a reports dashboard and then a purchasing module, the model interprets the intention and executes a sequence in real time:

1. Calls the get_top_selling_products tool against the sales API.
2. Calls the check_current_stock tool against the inventory microservice.
3. Crosses the data, calculates the shortage, and returns structured information.

All this processing happens under the hood, operating directly on the company’s pre-existing architecture and information. No line of business logic was rewritten; what already existed was exposed through a contract that the model can understand and execute.

The true optimization: The end of infinite views

The deepest technical impact of Zero UI is the drastic reduction of frontend development and maintenance cost. The previous use case, in a traditional paradigm, would have required programming a reports view, a dynamic table, date filters, confirmation modals, form validations, and a complex state manager to maintain consistency throughout the session.

Leveraging an agent scheme saves us from writing and maintaining millions of views and modals. It’s no longer necessary to pre-establish user behavior or code every possible branch in the application flow. The interface becomes ephemeral: it’s born to solve a specific intention and disappears when its purpose is fulfilled.

In a test environment built on Next.js and Vercel AI SDK, we explored how the interface mutates according to the backend response. The UI is assembled on the fly by returning React components directly from the server. If the execution of a tool requires confirmation to issue the purchase order, the system renders an actionable button; if the user needs to compare metrics, a table or chart is autogenerated at that moment. Visual components become ephemeral fragments tied to the conversation context, and are discarded when they’re no longer useful.

No more static screens waiting for users to discover them. No more menus nobody uses.

A new product vision

Designing product today requires changing perspective: the best graphical interface is the one that doesn’t exist until the exact moment it’s needed. Integrating the existing backend with models capable of orchestrating tools allows us to create infinitely more flexible systems. Eliminating the friction of pre-established interfaces not only transforms the user experience, making it natural and instant, but also radically optimizes the way we build and scale software.

The frontend of the future is not an application we download. It’s an orchestrator that listens, understands, and delegates.