What Is a Tool?
The word tool is used casually in everyday life — a hammer, a calculator, a search engine. In the world of AI agents, the word has a precise technical meaning. A tool is a defined, callable action that an agent can invoke by name, passing in specific inputs and receiving a structured output in return. It is not a vague capability; it is a contract with a name, a set of expected inputs, and a predictable output format.
The Anatomy of a Tool
Every tool in an AI agent system has at least three parts. The name is how the agent refers to the tool when it decides to use it. A name like web_search or get_weather is short, unambiguous, and describes the action. The input schema describes what information must be passed in. A weather tool might require a city name and a date. A search tool requires a query string. The schema is like a form the agent must fill out before the tool will run. The output is the structured result the tool returns. A weather tool returns temperature, conditions, and humidity — not a freeform paragraph, but organized data the agent can read and reason about.
A tool is a contract: give me these specific inputs and I will always perform this specific action and return this type of output. That predictability is what makes tools trustworthy.
How an Agent Calls a Tool
When an agent decides a tool is needed, it does not press a button or click a menu. It generates a structured message — sometimes called a tool call or function call — that names the tool and provides the required inputs. A system layer intercepts that message, runs the actual tool code, and sends the result back to the agent as a new message in its context. From the agent's perspective, calling a tool feels like reading a response from a very reliable assistant. The agent said: run web_search with query 'population of Brazil 2024'. The system ran it. The result came back: 216 million. The agent can now use that fact in its next thought.
An agent cannot actually reach out to the internet. It generates a structured request; a surrounding system runs the real code and feeds the result back. The agent sees only text in and text out — but the surrounding system has done something real.
Tools Are Not Abilities — They Are Interfaces
It is tempting to think of tools as superpowers the model has. That framing is misleading. The model itself has not changed — it still only processes and produces text. What has changed is that it is embedded in a system that intercepts certain outputs and acts on them. A tool is an interface between the agent's text-based thinking and the outside world's code and data. This distinction matters because it means anyone can build a new tool and give it to an agent. A school could build a tool that looks up a student's grade history. A hospital could build a tool that queries patient records. A business could build a tool that checks inventory. The agent's core intelligence stays the same; the tools expand what it can act on.
Complete the definition of a tool in an AI agent system.
Flashcards — click each card to reveal the answer
Which of the following best describes a tool in an AI agent system?
An agent calls a tool named get_stock_price and passes in the input ticker: 'AAPL'. What is the role of the input schema here?
Design a Tool
- Step 1: Choose a real-world capability you wish an AI assistant had — for example, looking up bus schedules, checking if a library book is available, or converting currencies.
- Step 2: Write a tool specification with three parts: (a) a name in snake_case like check_bus_schedule, (b) an input schema listing every piece of information the tool needs, and (c) an output description — what structured data does it return?
- Step 3: Write a sample tool call: pretend you are the agent and write out the message you would send to invoke your tool.
- Step 4: Write the sample output the tool would return.
- Step 5: Explain in two sentences why an agent using this tool would produce better results than a model answering purely from memory.