The Machine Ships the Pixels. You Still Own the Call.

Your team can now generate a full interface in seconds. So can your PM, your marketer, and the intern who watched one tutorial. That changes what your designers are actually for. A batch of new essays is trying to name the shift, and a few of them land something you can act on. Let me catch you up.

The work fell one layer down

John Maeda gave the move a name: UX to AX, user experience to agentic experience. The machine stops waiting to be operated and starts acting on what you meant. Adrian Levy takes that phrase, "moving from crafting interfaces to orchestrating outcomes," and points out where the design work actually went. Five manual steps collapse into one delegation. Everything that used to coordinate those steps drops below the screen, into a layer nobody was trained to draw.

So the pixels are cheap now. The coordination underneath them is not. When an agent takes an action on a user's behalf, someone has to decide what it is allowed to do, when it should stop, and what "done well" means. That is design work. It just does not look like a screen anymore.

Prediction is not the same as a decision

Here is the line worth stealing for your next review. A model learns what people prefer. It does not know why one choice is right. Aurélie Radom draws the split clean: preference describes what people choose, judgment explains why. A model can tell you clean layouts test well. It cannot tell you whether removing an element adds clarity or kills the personality that made the product work.

Her proof is the first iPhone. Every signal said serious phones needed keyboards. Business users wanted them. Makers competed on them. Removing the keyboard looked irrational through the lens of preference. Through judgment, it made something new. Optimization improves what exists. Judgment decides what should exist next. That second job is the one you defend.

Your org already tried to automate taste

Radom's sharpest point is that AI did not invent this problem. Committees have been replacing judgment with preference for years. A design proposal hits a meeting, and suddenly everyone is a reviewer. Product raises concerns, engineering names constraints, legal flags risk, sales speaks for the customer. Each voice is fair. The trouble starts when collaboration replaces ownership.

What comes out is not bad. It is usable, clean, consistent, and forgettable. "Average design is simply what emerges when every decision must satisfy every stakeholder equally," she writes. Healthy teams run on clear ownership, not consensus. Feedback should improve a decision, not replace the person making it. If your review process launders accountability across ten people, an agent will do the same thing faster, and you still get mush.

What happens when the thing works imperfectly

Takuma Kakehi has a small story that carries weight. He baked his daughter's birthday cake. The sponge came out flawed, the work of someone who had been baking for months, not years. He did not hide it. The whipped cream and strawberries made the whole thing work anyway. That is craft applied to something you cannot un-do.

That is the job now. Agents will hand your team output that is close but wrong in some way you cannot regenerate away. The design call is what to do with the flawed thing in front of you. Kakehi's line is the one to remember: the important decisions are not made when everything works, they are made when it does not.

The deep cut

There is a trust problem sitting underneath all of this, and it changes your roadmap, not just your headcount. Fabricio Teixeira points at the question users are starting to ask: "Wait, who made this?" That question barely existed a few years ago. Now, when anything can be generated, people are rethinking what they trust about creative work.

So the thing to protect is not the pixel-pushing. It is the named owner behind the call. When a review ends, someone should be able to say "I decided this, and here is why," and stand behind it if it fails. Assign that person before you hand more of the making to agents. An orchestration layer with no owner is not a strategy. It is a committee with better tooling.

Three questions for your team

• For our next three shipping decisions, who is the single named owner, and can they explain the why without pointing at a dashboard or a stakeholder vote?
• Where are we optimizing what already works when the real call is whether we are building the right thing at all?
• When an agent hands us output that is close but flawed, what is our process for deciding what to keep and what to override, and who makes that call?