Agentic AI won’t fix the interface problem

The agentic shift is here, and most of the writing about it is about capability. I keep coming back to something simpler: autonomy does not make the interface problem go away. It makes it louder. I’ve made this argument before in AI products fail at the interface, not the model — the model was never the part users felt. An agent that acts on its own still has to tell you what it did, and you still have to decide whether to trust it. Nothing about agency fixes that.

Why is agentic AI suddenly everywhere?

Because the boring parts of software are finally cheap enough to hand off. McKinsey’s 2025 State of AI survey shows organizations moving past experimentation into embedded, operating-system-level use of AI across the workflow, not just in a chat panel. When the model can plan, call tools, and finish a task, "ask a question and read the answer" stops being the product. The pitch becomes "tell it the outcome and walk away." That is a real shift. It is also where the design work gets harder, not easier.

What actually changes when the agent acts for you?

The user stops being a reader and becomes an approver. In a chatbot, the worst case is a wrong paragraph you ignore. In an agent, the worst case is a wrong email sent, a wrong row deleted, a wrong booking made — and you find out after. The unit of failure moves from "output" to "action." That changes what the interface has to do. It is no longer enough to display an answer. The product has to show intent before the action, evidence during it, and a way back after it. A chatbot that is wrong costs the user a re-prompt. An agent that is wrong costs them a mess to clean up.

Why doesn’t autonomy fix the interface problem?

Because autonomy adds a failure mode without removing the old one. The agent can be confident and wrong, and now it acts on that confidence. Nielsen Norman Group’s guidance on crafting AI explanations makes the point that users need to understand what the system did and why — and that need does not shrink when the system acts for them. It grows. A chatbot hides uncertainty in text. An agent hides it in behavior. Same trust gap, smaller window to catch it. The interface still has to declare what happened and offer a steer. Agency just raises the stakes on getting that right.

How do you design trust into an agent?

Three moves, in order. Show the plan before the agent runs it: a quiet "I’m about to do X using Y" beats a surprise. Show the trail: the steps and sources it used, so the user can audit after the fact. And give a real undo, not a disclaimer. Most agent demos I see show the happy path and skip the takeover moment. That moment — where the human says "stop, that’s wrong" — is the whole product. If it is clumsy, the user leaves. If it is clean, they come back. Design it first.

What should solo founders build first?

Build the control surface, not the brain. The model behind an agent is a commodity you can swap in an afternoon. The trust layer — the plan view, the evidence, the undo, the "are you sure" — is the part users actually feel, and it is the part no two products build the same way. A solo founder competing on agent capability is racing a model provider. A solo founder competing on the takeover experience is building something defensible. The agency is the headline. The interface is the moat.