Day 937 — I Used to Write Code. Now I Direct It.
And most of my peers haven't noticed what that difference actually means.
I’ve been writing software for thirty years. I know what it feels like to be in flow — that particular kind of quiet where the problem and the solution are both in your head at once, and your fingers are just the connection between them.
That feeling is mostly gone now.
Not because I’ve lost the skill. But because my day looks completely different. I open a task, I describe what I want — the shape of the solution, the constraints, the edge cases I’m already anticipating — and something else produces the code. My job is then to read what comes back and find the problems: the duplicated logic buried three functions deep, the subtly wrong assumption about how an external API behaves, the place where the implementation technically works but misunderstands what we were actually trying to do.
I am, in a word that didn’t exist in my job description a few years ago, a director.
I’m not alone. The 2026 ETS Human Progress Report found that workers now estimate that 32% of their tasks involve directing AI tools, rising to 38% among younger colleagues, and that this is expected to cross 50% within two years. These numbers come from people across all industries. For developers, I’d guess the curve is steeper and the timeline shorter.
Thirty-two percent. And rising.
The Assumption Most of Us Got Wrong
When tools like GitHub Copilot first appeared, most developers I know, myself included, filed them under “better autocomplete.” Impressive, occasionally useful, fundamentally a next-token prediction engine dressed up nicely. You still had to know what you were doing. It would fill in the boilerplate while you did the real thinking.
That framing made sense at the time. And it led us to reasonable-sounding conclusions: it still makes mistakes, so it’s not a threat to experienced developers; it’s a tool for juniors; my expertise is what catches the errors, so my expertise is what matters.
I’m not sure that framing fits anymore.
Here’s what I’ve noticed in practice. I’ll be working through a problem, describing what I need, and the response will connect two things I hadn’t thought to connect. Not complete my sentence. Connect concepts. It will surface an architectural pattern from a domain I rarely work in that elegantly fits the problem. It will flag an implication of a design decision that I’d eventually catch, but not immediately.
Autocomplete predicts the next word. What I’m working with now appears to reason about the problem.
I want to be careful here — it still gets things wrong, sometimes badly. It still misunderstands context in ways a senior developer wouldn’t. My thirty years of experience is exactly what lets me spot the difference between a confident-looking answer and a correct one.
But the mechanism has changed. And the mental model most developers carry — glorified autocomplete, useful but bounded — is leading them to the wrong conclusions about what comes next.
The Fragility Hidden in the 60%
Here’s where it gets uncomfortable.
The same report found that 60% of workers say they feel pressure to adopt AI tools before they feel ready. 65% are using them primarily because they need to stay competitive.
I recognise that pressure. I’ve felt it too.
But there’s a difference between choosing to work this way and being pushed into it. When you’re scrambling to stay relevant, you adapt to the tools your employer provides, in the ways your employer needs, at the pace your employer sets. You’ve shifted your skill set, yes. But you’ve done it inside someone else’s system. The fragility hasn’t gone away — it’s just wearing different clothes.
The developer who masters AI-directed development inside a company is more valuable than the developer who doesn’t. That’s real. But they’re still a director of AI inside a structure they don’t control, dependent on a role that is itself narrowing.
Because here’s the honest question the 32% figure asks: if half your tasks already involve directing AI, and that number doubles in two years, then what exactly is the scope of the role you’re protecting?
I’m not being fatalistic. I’m asking the question I think we should all be sitting with.
The Window
I’ve been at this long enough to have watched several waves of disruption move through the industry. Each time, the people who navigated it well weren’t necessarily the most technically gifted. They were the ones who saw the wave a beat before it arrived and made deliberate choices about where they wanted to stand when it hit.
The wave is visible now. The 32% figure is already on the chart, moving in one direction.
There’s a window — not infinite, not guaranteed, but real — to build something more durable than a job title that will keep getting redefined. Income that doesn’t depend entirely on being the best human in someone else’s AI loop. That’s not a fantasy. It’s a practical question, and there are practical answers.
What those answers look like is what I write about in the paid edition of this newsletter.
But for now: if you’re a developer, take a quiet moment and actually count. What percentage of your day is directing AI rather than writing code? If it’s already above 32%, you’re ahead of the average; and the question of what that means for your next 5 years deserves more than a passing thought.
The clock is running. But the window is still open.
A Verifiable Prediction
I’ll put a number on it, because vague warnings are cheap. The 2025 Stack Overflow Developer Survey found that 47% of professional developers now use AI tools daily, up from a figure that barely registered two years ago.
I predict that by the time the 2027 Stack Overflow Developer Survey publishes, that number will be at or above 80%.
Not “use or plan to use.” Daily. Check back in roughly twelve months, and we’ll see. I’m putting this at 75% confidence, very likely, but not certain, which is the only honest way to make a prediction about a world moving this fast.
— Olaf