So how do you actually use a tool that sounds equally sure when it is right and when it is inventing? Not by trusting everything, and not by trusting nothing. By turning a dial.

Picture trust as a gauge running from lean on it to check it carefully, and let the stakes set the needle. If a mistake would be cheap and obvious, brainstorming names, drafting a first pass, explaining an idea you can sanity-check yourself, leave the needle toward "lean on it" and move fast. If a mistake would be costly or hard to undo, a medical fact, a legal claim, a number going into a real decision, a citation you are about to repeat in public, turn the needle hard toward "check it," and verify against a source that is not the model.

That is the whole discipline, and notice it is not skepticism. A language model is not a truth machine; it is a very powerful pattern-completer with known, explainable failure modes. Being clear-eyed about exactly where those failures live is not distrust. It is what lets you trust it correctly, leaning where it is strong and checking where it is thin.

There is one capability left to understand, and it raises the stakes on everything in this chapter. So far the model only produces words. What happens when you let it take actions in the world, run code, send a message, change a file? That is what an agent is, and it is where this ends.