Tools: how an AI does things.

An LLM has no way to get new information unless you provide it. It can't search the internet, check your CRM, read your emails or look up today's weather just because you asked. Our picture of it is still: text in, text out.

So how does an AI app check the weather? The model gets help from the application around it. That application has a name, the harness, which comes up properly in the lesson on agents.

The application gives the model a description of each tool it may use: what the tool does and what inputs it needs. Say it describes a tool called get_weather that takes a location. You ask: "What's the weather in London?"

The model still can't check the weather. But it understands the question, sees that get_weather fits, and instead of answering, writes a structured request:

{ "name": "get_weather",
  "arguments": { "location": "London" } }

The important part: the model only asks. The application decides whether to do it, and does the actual work. That gap between asking and doing is where you put your safety checks.

Now try something more interesting: "Email Alice and ask if she's free tomorrow." The mechanism is the same. The risk is very different.

A

Read-only

Checks the weather, searches the web, reads a document. Worst case, you get a wrong or wasted answer.

B

Changes something

Sends an email, updates a record, books a meeting. Worst case, something happens in the real world that you can't take back.

For the email, a sensible application checks which Alice you mean, confirms her address, shows you the draft and asks before it sends. A tool that changes something outside the conversation deserves more care than one that only reads.

Tools let a model search the internet, connect to other applications, search document libraries, read and send email, create calendar appointments, update a CRM, run code and much more.

Tools aren't the only way to give a model information. You can paste text, upload a document or attach an image yourself. What tools add is that the model can fetch what it needs when it needs it, and act on your behalf. It starts taking part in the world outside the chat.