The 90-second gap: where your attention goes while the AI writes your code

If you code with an agent, you have a new kind of dead time: too short to start anything real, too long to just sit there. Mine is about ninety seconds. In that gap my hand goes to the phone before I decide anything, and by the time the diff is ready I have read four things I did not want to read. The fix that worked was not a better waiting routine. It was making the phone unavailable for the length of the session, using a physical gesture I have to repeat to undo.

The gap has a shape

Wait times with coding agents are not one number, which is exactly the problem. A rough map, from my own sessions (a guide on what to do while waiting for AI code assistants lands on similar numbers):

What you asked for	Typical wait
A simple refactor	around 30 seconds
A complex multi-file edit	around 5 minutes
An agent exploring the codebase	up to 20 minutes

Thirty seconds is nothing. Twenty minutes is a walk. The trouble is the band in between, and the fact that you cannot tell in advance which one you are in. You are not going to start reading a design doc for something that might resolve in forty seconds. So you do the one thing that fits any duration and requires zero setup: you pick up the phone.

That is not weakness. It is the only activity in the room that scales from ten seconds to ten minutes with no cost to start and no cost to stop. Your thumb has done the maths already.

The research watches your screen, not your hand

Here is where I have to be fair, because the data does not obviously agree with me.

Microsoft researchers studying how developers use AI agents found that developers followed the agent’s execution in real time for 84% of their prompts. They watched. Some deliberately avoided touching the codebase while the agent ran, so the diff would stay clean. That is the opposite of a distraction crisis.

But that study is measuring the computer. It tells you the developer did not switch to another repo or another window. It does not tell you what the other hand was doing.

METR ran into the same blind spot from a different direction. Their 2025 randomised trial famously found experienced open-source developers were 19% slower with AI tools than without, despite believing the opposite. When they tried to run it again, they hit a measurement problem worth reading twice: some developers could not report how long a task took, because they would “often work an unrelated task while waiting for the agent”. The waiting is now large enough, and fragmented enough, that it breaks time tracking.

So: you are looking at the terminal. You are also, some of the time, somewhere else. Both are true, and only one of them shows up in the logs.

The agent does not need a break. That is exactly why you have to take one on purpose.

Why this micro-wait is stickier than a normal break

Before agents, the waits in my day were compile time, test runs, and CI. Those were annoying but they were rhythmic. You learned their length. Agent waits are variable, and variable rewards are the whole design of a slot machine.

LeadDev reported in March 2026 on how compulsive this loop has become. They quote Steve Yegge saying the better you get at agentic coding, the more you want to use it. They cite a Multitudes study of more than 500 engineers finding a 19.6% rise in out-of-hours commits among developers using AI tools, and ActivTrak data showing productive hours up 46% on Saturdays and 58% on Sundays between 2023 and 2025. Developers told them they could not get up from the computer.

Built In followed in June 2026 with the cognitive side. Boston Consulting Group surveyed 1,500 workers: 14% reported a “mental hangover” from AI tool overuse, and those affected were 33% more likely to suffer decision fatigue. BCG also found productivity rising with the first, second and third agent, then declining with the fourth. Simon Willison is quoted saying he felt wiped out by 11 a.m. running four agents.

Put the two together. The loop already pushes you toward more sessions, later in the day, with less recovery. On top of that, every session hands you a dozen empty 90-second holes. Filling them with the most stimulating object you own is not neutral. It is what turns a tiring day into a fried one.

Which is why I made a choice that goes against most of what I see on my feed. Plenty of builders show off ten, twenty agents running at once. I went the other way, towards a kind of sobriety. One task, taken to the end. Very often one project, sometimes one corner of one project. And the 90 seconds are allowed to be what they are: a pause, a moment to think while it works, an idea that goes into a design file or onto paper. Not a slot to fill.

What I changed

Sobriety is one half of it: fewer things running, and a brain that gets to rest between prompts. The other half is what I did about the phone, because good intentions about the gap do not survive a phone within reach. I do not have a productivity system for that. I have a gesture.

When I start a real session with agents, I tap my phone on an NFC tag. Lock blocks the apps I chose, and there is no button on the screen that undoes it. Then the tag does not stay next to the laptop. It lives in the kitchen. If I want Instagram back, I stand up and walk there, which takes about eight seconds and almost never survives contact with actually standing up.

Three honest notes about this, because the article is useless if I oversell it:

• It does not stop you watching the terminal. Nothing here fixes the compulsion to babysit an agent. It only removes the second screen from the equation. That is a smaller claim than “fix your focus” and it is the one I can actually make.
• Whitelist before you regret it. If you are on call, or your team lives in Slack, block Instagram and TikTok and leave the work apps alone. A block that makes you miss a page gets uninstalled by Thursday. The support page covers the emergency unlocks and the session timeout, which exist for exactly this.
• The phone is not the only leak. Mine was, and sometimes still is, a browser tab. A phone lock does nothing about that. Two things I noticed, though. When the phone is out of the equation, I stay much more attentive to what is actually happening on my screen; a website can pull you in, but it does not trap you the way a phone does. And I started keeping a short list of small, comfortable sub-tasks I can do while the agent works: re-read the prompt, tidy a file, write the commit message, check a doc. Deciding those in advance means the gap already has something in it when it opens.

What surprised me was not the time saved. It was that the 90-second gap became usable. When the phone is not an option, the gap turns into re-reading the prompt I just wrote, which is where I catch about half of my bad instructions before the agent spends five minutes acting on them.

What to try this week

Nothing elaborate. One week, one change.

1. Name your gap. Next session, glance at the clock when you send a prompt and when you look back. You want the rough number, not a spreadsheet. Most people are surprised it is under two minutes.
2. Decide what the gap is for, once. Pick one thing: re-read the prompt, stretch, look out the window. Deciding in advance costs nothing. Deciding at second three costs everything, which is the point I made at more length in why willpower is not a feature.
3. Move the exit off the phone. Any tool that does this works. If you want the version I built, the setup guide takes about ten minutes and a €0.50 sticker. Put the tag somewhere that makes you stand up.
4. Use the gap to think about the change itself. Every time the agent is writing something, ask what need it answers, whether it was necessary, how it could go further. Kindly, not as a self-review. Combine that reflection with the code as it lands and the wait stops being boring. It becomes the part of the session where you actually think.

The agents are going to keep getting faster, and the gaps are going to keep getting shorter and more frequent. That does not make them easier to ignore. A gap you cannot fill is a gap you eventually use.