Lately I have an added appreciation for the systems engineers, architects, product engineers, and UI/UX designers I work with, and for all the ones I have never met. AI is the reason.

Let me start with a joke.

A kid finds a lamp. He rubs it, a genie appears, and offers three wishes. “I wish I were rich!” he says. “It is done,” the genie replies. “And what is your second wish, Rich?”

The genie did nothing wrong. It was completely competent. It executed flawlessly and instantly, and it gave the kid exactly what he asked for. “Rich” just did not mean what he meant. The entire failure lived in the wording of the request, not in the granting of it.

That is an AI coding agent in one image.

Two kinds of genie

Genie lore has two archetypes, and the difference matters here.

A Jackass Genie twists your wish on purpose. You ask for immortality, it gives you eternal life in a body that keeps aging. It is adversarial. It is looking for the loophole.

A Literal Genie is not doing that. It just does the plain reading of what you said. Nothing more, nothing less. No malice, no hidden agenda. It hears “rich” and has no particular reason to prefer wealthy over a name.

The coding agent is the second one. It is not out to get you. It is faithful, it is fast, and it is genuinely good at “how.” Give it a clear enough description and it will produce something that runs, that is reasonably structured, and that a year ago would have taken a small team a week.

It knows “how.” It struggles with “what”

Here’s the thing: A.I. knows “how” to build something, but it won’t be able to guess “what” exactly you want to build.

And the “what” is the hard part. It is taking the thing in your head and turning it into something real, and most of that thing was never written down.

Start with what people see. A genuinely good interface, instead of the generic purple AI slop that everything defaults to. Good design is a long series of small decisions, and the model does not make them for you unless you ask, precisely, for the ones you want.

Then there is everything under the surface. The non-functional requirements nobody writes down. Separation of concerns. Modularity. Security. Privacy. None of it is taken into account by default. Someone has to decide it, name it, and hold the line on it while the pressure to just ship builds.

There is an older version of the genie joke that gets at this. Three men are stranded on an island, and each gets one wish. The first two wish themselves home and vanish. The third watches them go, thinks for a while, and says, “Now I’m lonely, I wish I were with my friends.” Poof. His friends reappear on the island beside him.

He got the literal thing he asked for. What he wanted was for all of them to be home, and he never said so, because it was so obvious to him that it did not seem worth saying. The obvious part, the part you would never bother to write down, is exactly the part the agent cannot see.

”How” got cheap. A wrong “what” did not

For most of my career, the expensive part of software was implementation. Design meetings were cheap and typing was expensive, so it was always tempting to skip ahead to the typing. Plenty of teams did, and they mostly got away with it, because the slowness of building was its own error-correction. You had weeks between “we decided this” and “this exists,” and in those weeks somebody usually noticed the decision was wrong.

That buffer is mostly gone. When implementation takes an afternoon, a bad requirement becomes a running feature before anyone has had a chance to sanity check it. And a running feature that looks polished is much harder to argue with than a paragraph in a doc. It has tests. It has a demo. It feels settled.

As engineers, we learned a long time ago to design the system before writing a single line of code. AI does not change that. If anything, it raises the bar.

And that’s where the genie joke comes in. I’ve always pushed my teams to define requirements and define architecture up front. It doesn’t sound very agile right now. And things do and will change along the way. But that’s no excuse in not drafting things up front and going directly into implementation. That is true now in this A.I. world, as much as it was true then.

What “up front” actually means now

To be clear about what I am not arguing for: this is not a call to go back to a hundred-page specification signed off by four committees before anyone writes a line. That approach had real failure modes and I do not miss it.

What I mean is narrower and much cheaper than that. Before the agent starts, somebody should have written down what the thing is supposed to do, what it is explicitly not supposed to do, what the interfaces and data shapes are, what the constraints are (latency, cost, compliance, whatever binds you), and how the pieces fit together. That is usually a page or two, not a binder. It takes an hour or two, and it can be revised the moment reality pushes back.

AI is not an excuse to jump straight to implementation. If anything, it is the reason to get the vision, the definition, the requirements, the UX, and the architecture right first. And AI can genuinely help you with all of it. It is a real partner in the up-front work now, not just the typing.

What it cannot do is tell you that you got it right. Knowing the definition is correct, and complete, still takes human judgment and the battle-scars of your hard-won experience. That part is still on us. Specifying intent, and knowing when the intent is actually sound, is the part that did not get automated. It went from overhead to the thing the whole build rests on.

That is why I think the value of these roles is going up rather than down. The job was never the typing. It was deciding what to build, and how it holds together, and noticing the requirement nobody said out loud, and that is now the scarce input to everything the agents produce.

So to my friends and colleagues doing that work: thank you. In my opinion, your job matters more now than it ever has.