Primer

Volume I · 06 · Tools

06 · The runtime

Tools

A model cannot search, calculate, or edit a file. It can only emit structured calls. Something else has to run them.

In English

The model cannot search the web or run code. It can only write a structured request — “call calculator with 2+2.” Other software runs that request and pastes the result back. Then the model reads the result like any other text.

Try this
Step the call. Watch emit → execute → observe. That third box is not the model thinking. It is the world answering.
Keep this
Tools are hands. The model only names the hand; the harness moves it.

Lab · tool call

User

The question

Will I need an umbrella in Tokyo tomorrow?

Tool use is a protocol, not a capability of the weights. The model is trained to emit a reserved pattern — a JSON object, an XML tag — when a schema in context matches the job. The harness parses that pattern, runs code, and feeds the result back as more tokens.

If the harness is missing, “I called the weather API” is a story the model is telling. Nothing moved on the network.