Will agentic interfaces replace traditional UI?

It's difficult to predict the future of user-interface paradigms. Fifteen years ago, many confidently claimed that smartphones and mobile apps would replace desktop and web platforms. Mobile is hugely important; however, many successful companies still largely ignore mobile. Others argued that digital calendars, like Google Calendar and Microsoft Outlook, would render paper calendars obsolete, yet every Christmas, I still receive a Beaches of Hawaii wall-hanging calendar from my in-laws. The persistence of low-tech habits and legacy formats is a reminder that user preferences and institutional inertia matter. Don't assume agentic (AI-driven, autonomous) interfaces will replace traditional UIs.

Now, I hear predictions that, within a few years, all of our computer interactions will take place through an agentic interface. Will agentic interfaces actually supplant traditional UIs this time?

Why this prediction might come true

Agentic interfaces are getting a lot more useful. With Generative UI, agents now have the ability to dynamically generate just about any interface element on demand. That means agents will be able not only to fetch and visualize information but also to create inputs users can interact with to accomplish complex tasks. This capability removes the need for designers to manually design every conceivable screen and workflow.

Problems with Agentic Interfaces

There’s no doubt agentic interfaces will play a big role in the future of software. When it comes to usability, these interfaces fall short in important ways.

1. Lack of Visibility

When I say “well-designed,” I really mean “easy to use.” Folks who are bullish on AI often see natural language interfaces as simple because they hide all the system’s complexity behind a chat with an agent. But here’s where things start to get tricky. Sometimes agentic interfaces hide too much, making it hard for users to leave users guessing about what the system does.

A good tool doesn’t just work, it lets you see how it works. It gives you feedback about what’s happening and drops hints about how to use it, all without a manual. Think about doors: a push plate means “push,” a handle means “pull.” If there’s nothing there, you’re left wondering what to do next.

Let’s take elevators as another example. Imagine stepping into a voice-operated elevator. You say your floor, and it just takes you there. You look for the familiar buttons but find none... just blank walls. Suddenly, you’re left with questions: How many floors does this building have? Which floor am I on now? Is this even an elevator? Sure, after some trial and error with the voice agent, you’ll figure it out. You’ll probably miss having those buttons—not just the physical feel, but the instant feedback and control they offer. You’ll probably miss having those buttons—not just the tactile feedback, but also the instant feedback and control they offer.

In short, agentic interfaces can make things harder by hiding too much. If users can't see system status or available options, the interface becomes frustrates users and hinders task completion.

2. Inconsistency and Recall

Imagine you visit an app for your local zoo. It has data on all the animals who live there. When you last used the app you saw a really neat interactive graph of elephant births over time. Maybe you want to show it to your kid this time. "How did I get to the graph again?" You ask yourself. Usually, you'd look for clues like familiar phrases, highlighted links, open accordions, and pages you've visited.

In the agentic UI, these clues are missing. Your only option is to ask the agent. And unfortunately, LLMs are non-deterministic. It might assemble you something entirely new this time out of the lego blocks it's developers have given it to answer your questions. Each time you use the app, the outcome can vary because the underlying model is non-deterministic.

3. Unconventional

How does anyone know how to use anything? By looking at it and determining what they know about it from past experiences. Currently, agentic interfaces are new. So by nature they are unconventional. Some predict this will change as AI adoption becomes more widespread. I'm skeptical: these interfaces suffer from too many design problems.

My advice: don't rush to an agentic interface for your application unless your user research and testing indicates its the best way to solve your users' problems. Prepare for change and learn how these tools can benefit your application, but don't fall victim to a fad. When that convention becomes the default users expect, then you should certainly adopt it.

4. Natural language interfaces aren't always the best experience

Going back to the elevator example, it's one of those cases where a single button push is less effort than forming a sentence to express your desire. This is true with web experiences as well. If it's one or two clicks to get to your dashboard, that's less effort than saying 'show me my XYZ report' and waiting for a response, then going back and forth with the agent until the agent shows you the report.

The reality is, people don't like to type that much. This detail tends to get lost on software engineers, because we type all day as part of our jobs. We have keyboards that make delightful mechanical clicking sounds, and live by writing code. However, most people doesn't want to type that much.

"But you can use your voice instead," some of you might say. But that's another thing people hate doing. We want to interact in silence. We don't like to announce our intentions. Nobody wants their subway neighbor to hear them say "Claude, email my doctor for a refill on my hemorrhoid cream."

----

There's really no doubt that agentic interfaces have a role to play in our future and are here to stay, but much like mobile interfaces, they're a new way to interact with software. They're not a permanent replacement for all interfaces. I don't see sufficient evidence that agentic interfaces will become the default way people interact with software. Let's check back in a couple of years to see if it's a #agedlikemilk moment.