The same question answered with one model call and with a loop of two calls, showing that only the prompt changes.

The only new idea is a loop

A language model does one thing: read text, write text. It cannot search, click, or check. But the program around it — the harness — can agree to read one shape of text as a command, run it, and paste the result back.

That is the whole trick. Nothing inside the model became able to act. The search_query(…) it wrote is just text — text that something else agreed to obey.