When I was a student, I saw the movie Swordfish for the first time, and I laughed at the scene where Hugh Jackman's character, a hacker, writes a computer virus inside what looks like a 3D modeling tool. It seemed absurd. I had no idea how close to reality that absurd image would get.
Over the last few days, I built a few of my own pet projects with Claude Code, running several AI agents in parallel from one screen. That old Swordfish scene came back to me: a setup that looks almost absurd, and is genuinely powerful. In two days of part-time work, I produced more than I would have in a full week of full-time work a few years ago.
This isn't my day job. I've spent the last 20 years managing software teams, not writing code daily, so I have a good sense of how long this work normally takes. The feeling of directing five developers in parallel, except they answer in minutes, is hard to describe. And it left me with a question I can't stop thinking about: what is a programmer becoming?
The hard part was never the code
We used to think of programming as something almost mystical: a complex craft open only to a few people with rare knowledge. For years that held. The software engineer was, in a sense, an elite, treated almost like a god. That era is ending. A large part of coding is being commoditized; turning an idea into working code now costs far less, for individuals and businesses alike, than it used to. It's tempting to conclude that almost anyone can now build whatever they need.
But writing code was never the same thing as building something useful, and AI makes that gap impossible to ignore. Do we actually know what we need? Do we have enough good ideas? Do we know which tool would make the business run faster? That is the hard part, and no model hands it to you.
From coder to shaper
So the programmer is not disappearing, and I don't think the profession simply splits in two. The role is evolving. (A small minority will go the other way, deeper into the machine, becoming the computer scientists and researchers who need to know how everything works under the hood. But for most people in software, the change runs in the opposite direction.)
What is the programmer evolving into? Less a person who writes code, and more a person who shapes it.
Donald Knuth called his life's work The Art of Computer Programming. He meant "art" in the old, rigorous sense: a discipline. But programming is becoming an art in the other sense too. The programmer looks less like a typist and more like a sculptor. AI produces a rough shape from an idea, and the human works it until it fits the real need. Less time on the exact implementation, more on deciding what the thing should be, directing the work, and judging the result.
Because what always mattered was never memorizing internals. The valuable skill is no longer how to implement a piece of logic, but how to define what it should do: understanding the context, setting the direction, and mapping the path to the result. What does that take in practice? I'll come back to it. But first, the objection I keep hearing.
"But won't AI do all of it?"
The CEOs of the big AI labs are happy to tell you the profession is finished. In early 2025, Anthropic's Dario Amodei predicted AI would write 90% of code within months, and essentially all of it within a year. By some accounts, inside these companies, that has already happened. And with both labs now heading for the public markets, they have every reason to keep that story loud.
Maybe it's true. But it measures the wrong thing. Writing code and building useful software were never the same, and the better machines get at the first, the more the value shifts to the second.
Take one of those pet projects: a small video-processing app. The AI wrote code quickly, but it could not build the whole pipeline on its own. It made wrong assumptions about how the stages fit together. So I broke the work into stages myself, the way I would for a team, guided it through each one, and checked the result by hand at every step. Done that way, it processed everything correctly. The lesson stuck with me: agents are still weakest exactly where checking their work is expensive, where it takes real effort to tell whether they actually did the job right. The faster they produce, the more that judgment becomes the bottleneck.
What shaping takes: SHAPE
So what does shaping actually involve? After this experiment, and 20 years of watching how teams really work, I keep coming back to five things. They happen to spell the word.
Structure. Breaking a problem into the right pieces and describing each one clearly enough that a person, or an agent, can run with it. We are all becoming architects.
Handoff. Deciding, continually, what to do yourself and what to hand to the AI, then coordinating it. A good manager does this with people; here you do it with machines. The programmer becomes, in part, an AI manager.
Accountability. When code is cheap to produce, the hard part moves to everything around it: is it secure, does it do what it should, can you trust what came back? The machine writes the code, but responsibility for the result still sits with a person. Someone has to catch the vulnerability and stand behind what ships.
People. A programmer is no longer a lone introvert solving a puzzle in silence. You work with and through other people: aligning on the goal, communicating clearly, leading and being led. The more the machine types, the more the human's job becomes working with other humans.
Expertise. You can't build something valuable without understanding the problem you're solving. Deep knowledge of a vertical, like healthtech, fintech, or edtech, will be a bigger differentiator than fluency in any programming language. The mechanics of software can be learned relatively quickly; deep domain expertise takes a career.
What this means for how we train people
This is also why I worry about how we train new engineers. When I applied to study computer science, the entrance bar was math and physics, and that foundation still matters. (There's an Isaac Asimov story, The Feeling of Power, where people outsource arithmetic to machines so completely that a man who relearns to multiply by hand is hailed as a genius. We shouldn't forget how the machine works underneath.) But if the job is becoming shaping, a curriculum that teaches mostly coding is preparing students for the one part AI already covers. We need them solving real problems, in real industry context, far earlier than we do now. Otherwise we keep graduating people whose main skill is the thing AI now does for almost nothing.
So let me end with a question, especially for those of you who hire, teach, or mentor new engineers: what should a software graduate be able to do today, beyond writing code, to build a real career over the next five years?
I'd genuinely like to read your answer in the comments.
First published on LinkedIn. Discussion on LinkedIn