An AI calling outside tools, like a calculator, search or code, to get things right.
Language models are built to continue text, so they're shaky at exact math and at anything new, like today's weather. The fix is to give them tools: a calculator, web search, a code runner, a calendar.
The model doesn't run the tool itself. It writes a request in a set format, like "call the calculator with these numbers"; the app runs it and hands back the result, which the model uses to finish its answer. After OpenAI launched function calling in 2023, this became a standard feature of major AI services.
The real skill is knowing when to reach for a tool. Meta's 2023 Toolformer study taught a model to decide that by itself.
Think of it like this
On a test where calculators are allowed, the smart move on a giant multiplication isn't mental math; it's grabbing the calculator. Tool use is an AI making that same call.
For example
Ask "What's 1,234 × 5,678?" and a good AI calls a calculator instead of guessing.
Common mix-ups
Mix-upAn AI that uses tools runs the calculator or search itself.
ActuallyWhat the AI produces is an order form, with the tool's name and the values to use filled into set boxes. Many apps let you click a label like "searched" or "ran code" to open that form and the result that came back, so open it when an answer looks doubtful.
Mix-upThe more tools you attach, the better — always.
ActuallyEach tool's instructions go into the context window too, so the more you attach, the more room they take and the easier it is to mix up similar tools. And more tools that fetch outside text means more doors for hidden commands to get in.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- What signals tell an AI it's time to reach for a tool?
- Activity
- Make ten problem cards (a big multiplication, today's weather, a word's meaning, a short poem, what weekday it'll be next week…) and have groups sort them into "just answer" and "needs a calculator, search or calendar". For each tool card, write one line in a set format (tool name + value) for the request the AI would send to the app.
Records about this
Source-checked moments from AI history.
Tool calling was promoted from prompt parsing to a structured JSON-schema interface; the de facto standard API for agent development
Learns on its own, through self-supervised learning, when to call APIs and with what arguments — the route of building tool use into the weights