Waiting is the new context switch

I've shipped more in the past few weeks than in any comparable stretch of my career. I can also remember less of it than usual, and I've been walking around in a low-grade fog that I originally put down to sleep.

It isn't sleep. I think I've worked out what it is, and it's structural rather than personal, which is the only reason I'm writing it down instead of just drinking more coffee.

The gap is ninety seconds long

The loop has a shape. I describe a task. Something goes away and works on it. It comes back somewhere between ninety seconds and five minutes later.

That interval is the problem, and it's the problem specifically because of how long it is. Ten seconds and I'd wait. Twenty minutes and I'd start something real. Ninety seconds is precisely long enough to feel wasteful and precisely short enough that starting anything substantial is stupid. So it gets filled with whatever is nearest. The other terminal. A message. A tab I opened four hours ago for reasons I can no longer reconstruct.

Then the work comes back and I review it.

Reviewing generated code is a demanding cognitive act. You have to hold the intent, the surrounding system, and the specific diff in your head simultaneously, and decide whether something subtly wrong is hiding in something that looks completely fine. It wants a loaded, settled mind.

I am arriving at it from a tab about nothing, roughly forty times a day.

Sophie Leroy's work on attention residue described this years before any of us had agents. When you switch away from a task, part of your attention stays behind, and you arrive at the next thing carrying a fragment of the last one. What's changed is the frequency. My old workflow generated a handful of real context switches a day. This one generates dozens, and it generates them by design, because waiting is a structural feature of delegating work rather than doing it.

It collides with the part that needs focus

A couple of days ago I wrote about the verification tax, which was an argument that production got roughly ten times cheaper while verification got no cheaper at all, so the bottleneck moved to how fast I can genuinely understand output.

I missed something there, and it makes the problem worse than I described it.

Verification requires exactly the mental state that the workflow producing the work destroys. Careful review needs sustained, uninterrupted attention with the problem loaded. The loop that generates the review queue chops attention into ninety-second pieces. The tool manufactures a large volume of work that can only be done well in a condition the tool actively prevents.

Addy Osmani put numbers on the underlying gap in an essay on comprehension debt back in March. Generation runs somewhere around 140 to 200 lines of meaningful code per minute against a focused human's 20 to 40, which is a five to sevenfold gap between how fast code appears and how fast a person keeps up with it. He also reports that developers using AI generation as their primary workflow scored 17% lower on code comprehension assessments than people writing more of it by hand.

Those numbers assume a focused human. Nobody accounted for the reviewer being interrupted every ninety seconds by the same system generating the queue.

I'm not the only one feeling it. Boston Consulting Group surveyed close to 1,500 workers and 14% reported a mental hangover from tools running past their cognitive capacity, which the researchers took to calling AI brain fry. In the same piece, Simon Willison describes managing four agents at once as mentally exhausting and says he's often wiped out by 11am. When somebody that far from AI skepticism reports being cooked before lunch, the effect is probably real.

I am never stuck anymore

The second thing I've noticed took longer to see because it looks like an improvement.

I don't get stuck. Not properly stuck, not the way I used to, where a problem would sit on me for forty minutes and I'd pace around the kitchen turning it over. That state was unpleasant enough that I spent twenty years trying to minimize it.

It was also doing work I never credited it for. Being stuck forces the problem into you. You end up knowing the shape of the thing from every angle because you had no choice but to sit inside it. All the deepest understanding I have of any system I've built came from a stretch of being unable to move.

Now I get unstuck in forty seconds and the forty minutes never happen. I get the answer and skip the process that would have made the answer mine. It feels efficient in the moment and it leaves nothing behind.

The MIT Media Lab ran an EEG study on people writing essays with and without an assistant. The headline was about brain activity. The detail that stuck with me is that the assisted group struggled to quote their own essays and reported the lowest sense of ownership over work they had finished minutes earlier. Fifty-four participants is a small study and it's taken methodological criticism, so I won't lean on it hard. But that specific detail describes something I recognize. I have shipped things this month that I would struggle to explain a week later without opening the file.

What I'm changing about the loop

The obvious conclusion is to use the tools less, and I think that's wrong. The gains are real and I have no interest in giving them back.

What I'm doing instead is treating attention as the scarce resource in the system and scheduling around it.

Dispatch in batches. Instead of one task at a time with a gap after each, I queue several, then walk away from the machine entirely. The gap is going to exist regardless. Filling it with a tab is worse than filling it with a walk, because a walk doesn't leave residue.

Review in one unbroken block. Everything that came back gets read in a single session with nothing else open, at a time of day when I'm sharp rather than whenever the queue happens to fill up. Verification is the constrained resource, so it gets the best hours instead of the leftover ones.

Keep one thing a week deliberately slow. I pick something I could delegate and write it by hand, mostly to confirm I still can. The first time I did this it was uncomfortable in a way I found informative.

Notice the ninety-second reach. That reflex toward the other window is the actual mechanism, and catching it in progress is most of the fight. I'm not great at this yet.

None of that is a productivity system and I'd be suspicious of anyone selling it as one. It's an attempt to stop optimizing the part that got cheap while the expensive part quietly falls apart.

The thing I'd want somebody to tell me a month ago is that the fog has a cause outside my own discipline. It's what a loop produces when it interrupts you every ninety seconds and then asks you to concentrate.