AI Coding Agents and Attention

An AI coding agent collapses the distance between wanting a change and watching it happen. What that collapse removes is friction—and friction, it turns out, was load-bearing. It kept a programmer’s attention on one problem long enough to understand it, and it capped how many problems a person could hold in mind at once. Take the friction away and the cap goes with it. Attention is now the scarce resource in this work, not execution, and nobody has built the habits that scarcity demands.

That is the real claim sitting inside a louder one, from David Wilson: that these agents are “a thermonuclear ADHD amplifier.” The amplifier half of that phrase holds up, because an amplifier doesn’t generate a signal, it scales whatever signal is already there. So the charge isn’t that agents are bad for attention in some uniform way. It’s that they take an attentional disposition you already have and run it at a gain you couldn’t otherwise reach. Someone whose baseline already chases novelty, drops the tedious, and struggles to hold one thread will find that gain dangerous. Someone with a steadier baseline will mostly find the tool fast. The metaphor predicts that split, which is why dismissing the whole technology as horrific undercuts the better idea sitting next to it.

The mechanism is straightforward. Writing code by hand forces a kind of rehearsal: you cannot type a function without thinking it through, and the slowness of typing doubles as the slowness of intention. Working memory holds only so many live threads at once—a handful, by the classic estimate, though later work has trimmed that number—so you could only run so many tasks because starting each one cost you something. An agent removes that cost. Dispatch a task and a gap opens while it runs: a few seconds, a minute, too short to do anything real and too long to do nothing, so the next dispatch rushes in to fill it. Add a second agent, a third, and fragmented attention stops being an accident and becomes the default way of working. The tool’s throughput has no ceiling. A person’s does.

The same pattern shows up wherever friction leaves a skill. Calculators took over arithmetic and dulled the number sense that came with it; turn-by-turn navigation did the same to the mental map. The labor and the cognition that rode along with it left together. Coding is no different: the slowness of writing by hand kept you inside one problem long enough to understand it.

But not all friction was worth keeping, and Wilson’s framing never sorts the kinds. Much of what an agent removes is pure toil—boilerplate, configuration, another trip to the same documentation page—friction that taught nothing and returned nothing. Strip that and the freed attention goes to the parts that do demand it. The honest question is which friction, and horrific can’t answer that. The same tool that fragments one person’s focus can consolidate another’s, by taking over the executive overhead that was breaking their concentration to begin with. The amplifier runs in both directions; the ADHD framing only narrates one of them.

Borrowing a clinical term to describe what is really a report of how the work now feels raises a further problem. ADHD amplifier reaches for the authority of neurology to back a claim neurology doesn’t make. And the prior state wasn’t some attentional paradise being despoiled—programmers already worked inside Slack, an open Stack Overflow tab, an eleven-window IDE. One agent holding a single stated goal to completion can consolidate as much attention as the multi-tab sprawl it replaces scatters. Whether the tool fragments or focuses turns less on the tool than on the workflow laid over it: watching one agent work is a choice, running five in parallel is a choice, and closing one loop before opening the next is available regardless of how fast the tool runs. Calling the effect thermonuclear locates in the tool a determinism that lives in the user. A deadline-pressed engineer with a fast tool will widen scope and split attention regardless of what the tool is called; speed enables the breadth, but the human reaches for it.

The tool didn’t take that discipline from you. It just stopped requiring it—which means, for the first time, you have to supply it yourself.

Source: David Wilson, “The solution might be cancelling my AI subscription,” thoughts.hmmz.org, May 31, 2026 (also linked the same day by Simon Willison on simonwillison.net).