What to do while your AI is thinking
Jun 11 2026 · 4 min read · #AI #Agent #Prompt
Last week I watched an agent work through an invoice match, and about forty seconds in it wrote “hold on, let me reconsider the tolerance.” I caught myself staring at the screen, waiting to see what it would do next.
This is a new thing we all do now and nobody has said anything about it.
Nobody wrote the SOP for this
The pause used to mean something was broken. The spinner spun, you were allowed to be annoyed, and eventually you called IT. Everyone knew their role.
Now the pause means the machine is being careful. You can’t be annoyed at careful. You also can’t do much with the time, because you don’t know if it’ll be twelve seconds or four minutes.
And you chose this. That’s the part that gets me. There was a dropdown. You picked the slow model on purpose because the slow one is the smart one, and now you’re sitting there in a committed relationship with a progress indicator.
Nobody has told us what we owe our employers during this window. No training module, no policy, no line in the SOP. So everyone improvised, and after a few months of watching colleagues (and myself, badly) I think we’ve sorted ourselves into types.
The types
The Spectator reads every token of the reasoning like it’s a thriller. Leans in at “wait, actually.” Feels quietly insulted when the model writes “the user’s request is ambiguous,” because whose fault do you think that is. This is me. I know it’s me. I have given the thinking box more sustained attention than I’ve given company townhalls.
The Tab Goblin sends the prompt and immediately opens a new tab. Then eleven more. Loses the original entirely. Finds it again four minutes later through alt-tab archaeology, where the answer has been sitting, finished, waiting, slightly cold.
The Portfolio Manager refuses to sit idle, so they fire up two more agents in parallel. One’s drafting an email. One’s summarising a call. The thing about the Portfolio Manager is they’ve stopped doing any work at all. They spend the day checking in on direct reports who may be hallucinating. They’ve accidentally become middle management and they don’t seem to have noticed yet.
The Quality Assurer doesn’t trust the machine, so while it runs they start doing the task themselves in Excel. Just in case. They finish around the same time. They compare. The AI’s version is better. They use the AI’s version, delete their own file, and mention this to nobody, ever. I have done this. More than once.
The Prompt Regretter realises at second twelve that they left out the one constraint that mattered. Kill the run and lose everything, or let it finish confidently wrong and pretend that was the plan? They usually let it finish. Then they spend the remaining eighty seconds writing the correction, which means they’ve spent the whole wait working on the thing they were waiting for.
The Micro-Manager types “you still there?” at the twenty second mark. That’s it.
The parasocial soul
The reasoning stream does something strange to you over time.
When it says “hmm, this is tricky,” you feel vindicated. The task was hard. You were right to hand it over. Then it solves the whole thing eight seconds later and you feel worse than before.
When it guesses what you meant and gets it right, there’s a small warm thing that happens, which is an odd amount of feeling to have about a text box. When it guesses wrong you scroll up to check whether your prompt really was that vague. It was.
Fine, what should you actually do
Draft the next prompt while this one runs. Most people just wait.
Read the last output. Properly. Not the skim-and-“yep, looks right” thing we all do. The weak point in most AI workflows isn’t the model, it’s the review, and the review keeps losing to a scroll through your FYP. I say this as someone whose review has lost to Reddit this week.
Answer the Teams message you’ve been avoiding since Tuesday. It takes forty seconds and you have forty seconds.
Or ask whether the task needed doing at all. Be careful with that one. It ends careers and starts consultancies.
Honestly though, most of the value in the wait is that it’s the only gap left in the day where nothing is scheduled. Meetings took the rest years ago, then this thing turned up and handed back ninety seconds with nothing in them. Which we mostly fill with brainrot, so, you know. Progress.
Anyway
The agent finished while I was drafting this. It got the numbers wrong. Very thorough about it, though. Ninety seconds of visible reasoning, and a confident little summary at the end.
I’ve sent the correction. Back to doom scrolling.