The leverage tax

The pattern goes like this. I open the laptop and there are three or four threads in flight. A search relevance problem at work. A refactor on this site. A deploy script someone asked me to look at. A draft talk outline. Each one has an agent attached. To keep the agents moving, I have to reload each context back into my own head first: where did I leave this one, what did the last run output, what is the next reasonable instruction. Then I fire the instruction, switch tabs, and start the ritual again on the next thread.

As a lead before AI I was already running at something like 10x. Now the outputs look more like 100x, and on a good day 1000x. And I am not particularly happy about it.

Something is being borrowed

The leverage is real. I am not going to pretend it is not. What has changed is the shape of the work. The thinking I used to do inside a problem, the slow turning-it-over that happened while I typed, now happens in short bursts between agent runs. I still make the calls, but I make them on a thinner slice of understanding than I used to.

There is a paper out of MIT Media Lab from mid-2025, Kosmyna et al.’s Your Brain on ChatGPT, that puts language around this. It is a small EEG study, still a preprint, so worth reading as suggestive rather than settled. The framing the authors land on is “cognitive debt”. When you offload part of a task to an assistant, your own networks stay under-engaged, and that under-engagement persists into the next task, even when the assistant is gone.

Addy Osmani writes about the day-to-day version of the same thing in Your parallel agent limit. He names two costs I recognise in myself. An “ambient anxiety tax” from threads I have not checked in a while, and “comprehension debt” from output the agent produced faster than I could read it properly. Every thread I add does not cost nothing. It costs a bit of the attention I was using to hold the other ones.

Quarterly wins, five-year costs

Companies book this output. They book it this quarter, in shipped features and closed tickets and reduced cycle time. That is a rational thing for them to do. The harder question is who pays for the difference between the output and the capability behind it.

METR ran a randomised study on experienced open-source developers in early 2025. The headline finding was that the AI-assisted group was about 19% slower, while the developers themselves expected to be 24% faster and, after the fact, still rated themselves 20% faster. A 39-point gap between what was actually happening and what it felt like. That is the kind of perception error that is very hard to self-correct, because your own sense of whether you are getting sharper or shallower becomes unreliable. It compounds quietly.

The question I keep sitting with is a five-year one. If the engineer who ships at 100x in 2026 is a meaningfully shallower version of themselves in 2031, who carries that cost? Not the company; companies are fine with the trade if the output keeps arriving. The engineer carries it. So does the cohort of juniors who are trying to build the deeper version of the craft in the first place, on a track where most of the easy problems already have agents on them.

It is a variation of something I wrote about in Stoicism and the critical path: diligence on the wrong problem is still irrelevance. Output at 100x on work that does not build the capability you actually need is the same shape of mistake, just dressed up better.

What I am changing

A few small things, all of them deliberate:

• Cap parallel agents. For anything genuinely hard, one serious thread at a time. Four concurrent ones is a management job, and managing is not the same thing as engineering.
• Pick a few problems each week and solve them without AI, end to end. Not as a statement, just as a way of noticing which muscles have gone soft.
• Write the design before the agent does. If I cannot explain the shape of the solution in a paragraph, the agent should not be touching it yet.
• Treat review as real work, not a rubber stamp. Read the diff the way I would read it if I had written it, because in the way that matters, I did.
• Keep writing notes like this one. Unloading the thinking from my head into prose is how I check whether the reasoning was mine or borrowed.

Leverage without the underlying craft is decoration. I would rather be a bit slower, and still be the engineer who could have done it without help.