In one sentence
A tool is a defined action an AI is allowed to perform, such as looking up a record or creating a ticket.
The simple model
On its own, a model produces text. A tool lets it do something in the world: search a knowledge base, read a customer record, create a ticket, send a message. Each tool is a specific, defined capability you deliberately give the AI.
The difference is like an assistant saying "you should create a calendar event" versus being given a button they are allowed to press to create one. The tool is that button.
Why you care
Tools are exactly where an AI can begin affecting external systems, which is why each one is a permission decision, not just a technical detail.
- Every tool you add widens what the AI can do, and what it can get wrong.
- A tool that only reads is far safer than one that changes data. The read versus write distinction applies to each tool.
- Tools are usually built on APIs: the tool is the specific action, the API is how it reaches the system.
Giving an agent a tool is giving it authority. Give the fewest tools that let it do its job.
Design tools narrowly
A narrow tool is safer and easier to reason about than a broad one. "Create a ticket in the support queue" is safer than "run any command on the ticket system." Prefer tools that do one defined thing, with clear limits on what they accept.
A small example
An agent needed to help with support. It was given three tools: search the knowledge base (read), read a ticket (read), and draft a reply for approval (no direct send). It was deliberately not given a tool to close tickets or email customers. Its capabilities were exactly its permissions, and nothing more.
What this does not solve
Defining tools sets what the AI can do; it does not guarantee it uses them correctly. A tool can be called with wrong inputs, or at the wrong time. Validation, limits and, for consequential tools, approval are still needed around each one.
