What is tool use in LLMs, and why is it essential for building agents?

What you need to know

The three limits tools remove

What a tool call looks like

In the Anthropic API, a tool is declared like this:

1{2  "name": "get_balance",3  "description": "Get the current balance of one of the user's accounts. Use when the user asks how much money they have.",4  "input_schema": {5    "type": "object",6    "properties": {"account_id": {"type": "string", "description": "e.g. SB-0042"}},7    "required": ["account_id"]8  }9}

The model replies with a content block such as {"type": "tool_use", "id": "toolu_01…", "name": "get_balance", "input": {"account_id": "SB-0042"}}, and the response has stop_reason: "tool_use". OpenAI's Responses API uses {"type": "function", "name", "description", "parameters"} for the definition and returns a function_call item whose arguments is a JSON string you must parse.

Why it makes agents safe enough to ship

A tool call is a clean boundary. Before running it, your code can:

• Validate the arguments against the schema and business rules.
• Authorise it for this user ("can this customer see this account?").
• Gate it behind human approval if it has side effects.
• Log it for audit and debugging.

Free text gives you none of that. "I'll transfer ₹5,000 now" in prose is impossible to police; a create_transfer call with typed arguments is easy.

How the model learned it

Current models are trained on tool-calling data, so they choose tools from the description and fill arguments from the schema. That is why the description is so important: it is the only information the model has when deciding.

A real-life example

A bank's assistant has three tools: get_balance, list_transactions and create_transfer. A customer asks, "How much did I spend on Swiggy last month, and can I afford to send ₹20,000 to my sister?"

The model calls list_transactions(account_id="SB-0042", merchant="Swiggy", from="2026-08-01", to="2026-08-31") and get_balance(account_id="SB-0042") in the same turn. The results come back — ₹6,480 across 14 orders, and a balance of ₹48,200 — and the model answers with exact numbers. It then proposes create_transfer(amount=20000, payee="Priya"). The harness sees this tool is marked "needs confirmation", shows the customer a card ("Send ₹20,000 to Priya S., UPI priya@okbank?"), and runs it only after the customer taps Confirm.

Without tools, the model could only guess a balance — which in banking is worse than useless.

Follow-up questions to expect

• "Is tool use the same as structured output?" — Related but different. Structured output constrains the final answer to a schema; tool use asks your code to do something and continues the conversation with the result.
• "What if the model calls a tool that doesn't exist?" — Your dispatcher should return an error result naming the valid tools, not crash.
• "Can the model call several tools at once?" — Yes, most APIs allow parallel calls in one turn; run them concurrently and return all results together.