The description of a tool that the model reads in order to decide whether and how to use it. Name, what it does, what arguments it takes. It is a prompt, not just a technical interface, and writing it well makes more difference to whether tools get used correctly than almost anything else.
Technical view
Typically JSON Schema with a natural language description per function and per parameter, injected into context on every request and therefore counting against the token budget. Selection accuracy degrades as the number of tools grows, which motivates dynamic exposure of a relevant subset. Descriptions should state when not to use the tool as well as when to, and they are untrusted-adjacent, since third-party schemas can carry injected instructions.
Related
Part of AI Terminology Index, section: Agents and Tools.