Last week, I watched one of our students describe an idea out loud, hand it to a coding agent, and walk to get a coffee. By the time she came back, the thing existed. Not a mockup, not a Figma flow with hotspots taped over it, but a working prototype she could click, break, and argue with. She looked at it for a second and said, almost disappointed, “That was too easy.”
I have spent years arguing that the most interesting professional of this decade is the one who can close the gap between imagining something and actually making it. That was the whole premise. The handoff between the designer who dreams and the engineer who builds is where intention goes to die, and the person who can hold both sides at once is rare and valuable.
What I did not fully reckon with is what happens when the gap closes on its own. When the building part, the thing that used to take the skill, becomes the part you can walk away from while the coffee brews.
Two writers I follow have been circling this from opposite ends, and reading them together this month rearranged something for me.
From working with AI to managing it
Ethan Mollick put a name to the shift. In The Shape of the Thing, he argues we have crossed out of the era of prompting AI back and forth and into something else entirely.
“Starting in late 2025, we entered a new era thanks to AI agents... These are AI systems that you can just give work to, sometimes hours of human work, and get back reasonable and useful results in minutes. This is an era of managing AIs, rather than working with them.”
His example is almost comically blunt. A three-person team at StrongDM built what they call a Software Factory, running on two rules: code must not be written by humans, and code must not be reviewed by humans. The agents write it, other agents test it inside a simulated customer environment they build for themselves, and the humans only see the finished product before it ships.
Read that as a designer, and you feel the floor move. The part of my job that was hardest to fake, the actual construction, the thing that separated people who could build from people who could only describe, is being quietly automated. The handoff I spent my career trying to eliminate is being eliminated, just not by me.
The leverage is real, the floor has risen
Linas Beliūnas is looking at the same shift through the lens of money and startups, and he is more excited about it. In The One-Person Unicorn, he describes building an entire startup operating system out of Claude skills, twelve interconnected playbooks covering everything from idea validation to managing a board, then running a fictional company through the whole thing. His conclusion is the line I keep returning to:
“AI outputs are only as good as the architecture and data behind them. Change the architecture, and the outputs change completely.”
In a companion piece, What to Build in 2026, he reads the venture capital consensus and finds every major firm saying the same thing: the language has shifted from augment to replace. The pitch is no longer software that helps the lawyer. It is software that is the lawyer, for a well-defined scope. He calls the model Service-as-Software, and notes it sells into the services budget, which dwarfs the software budget by an order of magnitude.
Put Linas next to Mollick, and you get the full picture. Mollick describes the capability, you can now hand off hours of work. Linas describes the consequence: the minimum viable team is shrinking, and the floor of what one person can produce has risen dramatically.
So what is actually scarce now?
Here is where I land, and it is not where I expected.
If building is no longer the bottleneck, then building was never the point. It was a proxy. We valued the person who could build because building was hard and expensive, and the ability to do it was a reasonable signal that someone had taste, judgment, and the stubbornness to see an idea through friction. The skill was a stand-in for something underneath it.
The agents take the skill and leave something underneath exposed. Linas says it plainly: the output is only as good as the architecture behind it, and architecture is just another word for decisions made on purpose. Mollick, in his other recent piece, Choosing to Stay Human, warns about the opposite, what his Wharton colleagues call “cognitive surrender,” the very human habit of taking the AI’s confident answer and stopping there.
“Agentic systems are designed to make your life easier, because they just do stuff. Which is great for getting stuff done, bad for learning anything, or staying authentic, or avoiding cognitive surrender.”
This is the design engineer’s specific risk, and it is not the risk most people are worried about. The danger is not that the agent cannot build. It is that it builds so smoothly, you stop noticing what it decides on your behalf. Every default it picks is a design decision. The empty state it generated, the error message it wrote, and the way it handled the user who does not behave the way the happy path expects. The agent has opinions about all of these, and if you are getting coffee, they become your product’s opinions by accident.
The student who said “that was too easy” was right to be uneasy. The work did not disappear. It moved. It moved from typing the thing into deciding whether the thing it typed is any good, for whom, and at what cost to the person on the other side of the screen.
That decision is not faster than it used to be. If anything it is slower, because nothing in the tool forces you to make it. The friction that used to be in the building, where you could not avoid confronting your choices, has to be put back deliberately, somewhere else, by someone who knows it matters.
I think that is the actual job now. Not closing the gap between imagining and building, the machines closed it. The job is keeping the human decision in the loop after the gap is gone, choosing where to slow down on purpose in a process designed to never let you. The leverage Linas describes is real, but it amplifies whatever taste you bring, including the absence of it. Ten times the output is ten times the quality or ten times the slop, and the tool is genuinely indifferent to which.
The design engineers I would hire tomorrow are not the ones who prompt the fastest. They are the ones who look at what the agent produced in ninety seconds and can tell you, specifically, what is wrong with it.
So here is what I am still chewing on, and I would like to know what you think. If building is free, the scarce thing is knowing what is worth building and what “good” actually means for the person using it. But that judgment used to be earned the slow way, by building badly for years until you developed a feel for it. If the next generation never builds badly, because the agent never lets them, where does the taste come from?
Sources
Ethan Mollick, The Shape of the Thing (One Useful Thing, March 12, 2026): https://www.oneusefulthing.org/p/the-shape-of-the-thing
Ethan Mollick, Choosing to Stay Human (One Useful Thing, May 26, 2026): https://www.oneusefulthing.org/p/choosing-to-stay-human
Linas Beliūnas, The One-Person Unicorn (Linas’s Newsletter, February 2, 2026): https://linas.substack.com/p/onepersonunicorn
Linas Beliūnas, What to Build in 2026 (Linas’s Newsletter, February 12, 2026): https://linas.substack.com/p/what-to-build-in-2026