These days I hand a lot of work to AI agents. Write the code, dig up the sources, draft the thing — they’re fast, and they’re good at it. I’m clearly getting through far more than I used to.
And then something odd dawned on me. My throughput had gone up. The reach of my own thinking hadn’t moved an inch.
Scene one. I give an agent a nasty bug, and a few minutes later it hands back a fix. The code runs, clean. But when someone asks me why it was broken and why this fix is the right one, I can’t really answer. The output is in my hands; the understanding that holds it up isn’t in my head. The work got done. I just never understood it.
Scene two. Before a talk, I once handed off the research wholesale. A tidy document came back, and I built the slides straight from it. The presentation went smoothly — right up until someone asked, “That figure — how does it compare to two years ago?” One layer down, a single question, and I stalled. The surface was flawless. Underneath it, nothing.
Between the two scenes you can see the line: what gets delegated and what doesn’t. What’s delegated is the execution of thinking — the labor of gathering information, writing the code, producing the sentences. What isn’t is the understanding that holds the work up: defining what the real problem is, weighing why this path and not another, and finally owning the call. Hand off the first and a result still comes back. Hand off the second and something simply goes hollow inside you.
Why does only one side empty out? Because execution leaves a result behind, while understanding forms only in the doing. You stare at the bug yourself, hit a few dead ends, float a theory and watch it break, and only then comes the oh — so that’s what it was. Understanding is the residue that settles out of all that fumbling. When someone else does the wandering for you, only the conclusion crosses over, clean. The wandering itself never makes the trip. So you can hold the answer in your hand while the understanding that produced it stays outside you.
Here’s the obvious objection. So am I supposed to understand everything myself? We use compilers without knowing how they run, electricity without knowing why it flows, and we manage fine.
True — but that’s a different kind of not-knowing. A compiler is a verified, settled foundation. Plenty of people have checked what’s underneath it, and when it’s wrong, it shows fast. We stand on that solid layer and build our own understanding on top of it. Abstraction is stepping up one floor from a floor you trust.
Delegation turns dangerous when the direction flips — when what you empty out isn’t the floor you stand on but the very floor you’re responsible for. The figure I choked on wasn’t a foundation I was standing on; it was the thing I was supposed to be explaining. Abstraction leaves the understanding to the foundation and builds higher. Delegation gives away the whole spot where the understanding was supposed to sit. From the outside both look like “letting someone else handle it,” but one rises and the other caves in.
Two people, a century apart, put their finger on the same line.
“You can outsource your thinking, but you can’t outsource your understanding.” — a line of unknown origin, spread by the deep-learning researcher Andrej Karpathy, who likes to quote it
“Knowledge can be communicated, but not wisdom.” — Hermann Hesse, Siddhartha (1922)
Karpathy’s “execution of thinking” and Hesse’s “knowledge” are cut from the same cloth: both are things you can hand over. As work or as sentences, they travel. “Understanding” and “wisdom” don’t. They grow only inside the person who lived and worked their way through them. You can be told how to cross the river a hundred times; the feel of the current pushing back is the one thing no one can pass to you. It’s no accident that Siddhartha ends at a riverbank, with a man who gets there on his own — not because anyone taught him.
What’s changed across that century is one thing. The surface that can be handed over — the execution of thinking — has become incomparably cheaper and easier. In Hesse’s day, moving knowledge took books and teachers; not long ago, a line of code was still typed out by a person. Now you can offload that entire surface onto an agent. Which makes the illusion easy to fall for: the surface got this easy, so surely the depth came with it. But what doesn’t cross doesn’t cross, no matter how the tools change. If anything, the easier the surface gets, the easier it is to finish the job holding only that — the depth left empty.
So lately I stop for a second before handing work off. In this task — defining the problem, choosing the path, making the call — am I in it? Or am I about to take the result and leave the understanding empty? When the answer feels off, I slow down and work through it with the agent one step at a time. A trade: give up a little speed, keep the understanding on my side.
The tool carries the surface. The river, I still cross myself.
— tomte