Tools: An Agent's Hands
A language model by itself is like a brilliant mind trapped in a sealed room. It can think, reason, and communicate through a slot in the door — but it cannot reach out and touch anything. Tools are what break open that room. They are the agent's hands. When an agent has a tool like web search, it can actually retrieve real-time information. When it has a code-execution tool, it can run programs and see real output. When it has an email tool, it can actually send messages. Tools transform the agent from a text generator into a capable worker.
What a Tool Is, Technically
In technical terms, a tool is a function that the agent's language model is allowed to call. The agent's developers define each tool with three things: a name, a description of what it does, and a specification of what inputs it needs and what output it returns. The language model reads those descriptions and, when it needs to accomplish something that matches a tool's description, it generates a structured request to call that tool — like writing a command rather than prose. The agent framework intercepts that command, actually runs the function, and feeds the result back to the language model as new context. From the model's perspective, it thought about what it needed, asked for it, and received an answer.
Tool calling (sometimes called function calling) is the mechanism by which an agent's language model requests an external action. The model outputs a structured specification — which tool to call and with what arguments. The agent framework executes it and returns the result. The model never directly runs code; it asks for things to be run on its behalf.
A Catalog of Common Agent Tools
Agent developers can give agents almost any tool they can write a function for. Some tools appear in almost every practical agent because they address fundamental needs. Web search lets the agent find current information that was not in its training data. Given a query, the tool returns a list of search results and their content — real-time, from the live internet. Code execution lets the agent write a program and then actually run it, seeing the output. This is powerful for math, data analysis, and generating charts. File read/write lets the agent open documents, read their contents, and save new or modified files to a computer's filesystem. Calendar and email tools let the agent interact with scheduling systems and send or read messages — enabling genuine communication and scheduling on a user's behalf. Database query tools let the agent retrieve or update structured data, connecting it to records in spreadsheets, SQL databases, or cloud data stores. Browser control tools let the agent actually navigate websites, click links, fill forms, and extract information — allowing it to interact with any website, not just those with a formal API.
Match each tool type to the specific real-world capability it gives an agent.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
How an Agent Chooses Which Tool to Use
The agent does not randomly call tools. At each step of its reasoning loop, it considers its goal, its current state, and the tools available. It picks the tool whose description best matches what it needs right now. Good tool descriptions are crucial. If the description is vague, the model might call the wrong tool or miss the right one entirely. Designing clear, precise tool descriptions is one of the most important jobs an agent engineer has. An agent can also chain tools together. It might search the web to find a document, use a file tool to save that document, then use a code tool to extract the statistics it needs from the saved text, then use an email tool to send those statistics to a teammate. Each tool's output becomes context that informs the next tool call.
Every capability gap between a chatbot and an agent — live data, real files, actual messages, running programs — is closed by a tool. If you can write a function that does something, you can give that capability to an agent. This means agents are only as powerful as the tools their designers build for them.
An agent is writing a data analysis report and needs to calculate the average of 50 numbers accurately. Which tool is most appropriate?
What three things does an agent's developer specify when defining a tool?
Build an Agent's Toolbox
- Step 1: Pick one of these agent missions: (A) Help a student write and submit a science fair abstract, or (B) Automatically report the weather and traffic conditions to a commuter every morning.
- Step 2: List every tool your agent would need to complete the mission. For each tool, write: the tool's name, what it does in one sentence, and one example of how the agent would use it for this specific mission.
- Step 3: Identify the most critical tool — the one without which the mission would fail entirely.
- Step 4: Now imagine you are NOT allowed to use that critical tool. How would you redesign the agent's approach? Can it still complete the mission, or does it fall back to being just a chatbot?