What a Loop Means in AI

Nobody writes the good version first

Think about the last thing you wrote that you were genuinely pleased with. A proposal, a difficult message, a set of slides. You did not produce it in one motion, start to finish, correct on the first attempt. You wrote something rough. You read it back and winced at the second paragraph. You cut it, moved a section, checked a number, read it again, and stopped when it was good enough to send.

That is four moves repeated: try, check, adjust, repeat. It is so ordinary that we barely notice we are doing it. It is also, almost exactly, what a modern AI tool is doing when you watch it think, retry, and work through steps on screen.

One pass versus many

When you ask a chatbot to name the capital of Portugal, you get a single pass. The question goes in, an answer comes out, and nothing in between examined that answer. For a question like that, one pass is plenty.

Now ask the same tool to go through a file of last quarter's expenses, flag anything that looks miscategorised, and write you a summary. One pass cannot do that honestly. It would have to open the file, and it cannot know what is in the file until it looks. It would have to notice that some rows are formatted differently, which it cannot know until it hits them. Something has to happen, then be looked at, then change what happens next.

So the tool loops. It takes an action, reads the result, decides what that result means for the plan, and takes another action. Everything you see labelled as thinking, or as steps, is this cycle made visible.

YOU:  What is the capital of Portugal?
AI:   Lisbon.
Done. One attempt. Nothing was checked, and nothing needed to be.

YOU:  Go through this expense file and flag anything miscategorised.
PASS 1  Action: open the file
Result: hundreds of rows, and one column has mixed date formats
Adjust: tidy the dates before anything else
PASS 2  Action: tidy the dates
Result: a handful of rows fail, they contain notes not dates
Adjust: set those aside and continue with the rest
PASS 3  Action: compare each row's category against its description
Result: fourteen look wrong
Adjust: re-read those fourteen before reporting them
PASS 4  Action: re-read the fourteen
Result: eleven genuinely wrong, three were fine
Adjust: nothing left to adjust
STOP    Report the eleven, plus the rows it could not read.

Notice what pass 1 bought. Nothing that appeared in pass 2 could have been planned in advance, because nobody knew the dates were inconsistent until the file was open. That is the whole point. A loop lets the work respond to what it finds.

The moment the loop stops

A loop needs a way to end. Sometimes it is a finish line: the task is done, the checks pass, the document meets the brief. Sometimes it is a limit: enough attempts, enough time, enough money. Sometimes it is a question back to you.

Human loops end the same three ways. You stop editing the proposal when it is good enough, or when it is due, or when you send it to a colleague and ask what they think. An AI loop has all three of those exits available, and if you do not choose one, something else will choose for you.

TRY      do the next thing
CHECK    look at what actually happened
ADJUST   change the plan based on what you saw
REPEAT   until finished, out of budget, or stuck

Quick Quiz