Atlas · agents
Tool use
The model replies with a structured request naming a tool and its inputs; the harness decides whether to run it.
See it working: Plate I
In plain words
A model cannot open a file by itself. It can ask. A tool request is a short structured message that says which tool to use and with what inputs.
How it works
The harness lists the available tools, with names and descriptions, in the context. When the model wants one, its reply names the tool and gives arguments. Your application code executes it and hands the result back as the next piece of context.
Going deeper
Choosing a tool and executing it are separate steps owned by separate parties. That gap is where permissions, logging and sandboxes live. Formats differ between providers, but every one is a name plus structured arguments.
Sources
- Function calling OpenAI
Execute code on the application side with input from the tool call
- Model Context Protocol: Tools (2025-06-18) Model Context Protocol
Tools enable models to interact with external systems
Quotes were checked against the live pages on 2026-11-08.