What a loop is

A single model call takes a question and returns text. It cannot look anything up, so when it is asked about something it has no access to, it has two options: say it does not know, or invent an answer. Models are trained to be helpful, so they frequently invent.

An agent loop changes the shape of the interaction. Instead of one call, the model is called repeatedly, and on each call it may either finish or ask for something. When it asks, the caller runs the request, feeds the result back, and calls again. That repetition is the entire mechanism.

The lab beside this lesson runs the same task two ways. "No loop" sends one call with no tools, so the model answers from nothing. Any specific detail in that answer — a status, a date, a carrier — was invented, because the information was never available to it. "ReAct" gives it tools and lets it iterate.

Watch what ReAct does not fix, though. Giving the model tools makes grounding possible; it does not guarantee it. A model can call the lookup, read a status and a placed date, and still promise a delivery window the tool never returned — the fact that a tool ran is not the same as the answer using only what it returned. The lab runs a small grounding check on the reply and flags exactly this: a delivery estimate or a shipping date the observations never contained.