What is a tool? (Function calling)

By itself, an LLM is a very well-read person locked in a windowless room. It can reason, write, and remember things it saw during training — but it can't check today's weather, query your database, or send an email. Tools solve this. A tool is a function YOU provide that the model can decide to call. The model itself doesn't execute code — it outputs a structured request to call your function; your code runs it and returns the result.

You define each tool with: name (what to call it), description (what it does — this is the routing signal), and input_schema (a JSON Schema describing its parameters). The description is critical: the model reads it to decide when to use this tool. Vague descriptions = vague behavior.

{
  "name": "get_weather",
  "description": "Get the current weather for a city. Returns temperature in Celsius and a text summary.",
  "input_schema": {
    "type": "object",
    "properties": {
      "city": {"type": "string", "description": "City name, e.g., 'Toronto'"}
    },
    "required": ["city"]
  }
}

You send Claude a message with the tools available. If Claude decides to use one, its response contains a tool_use block with the tool name and the input. YOU execute the tool. YOU send back a new user message containing a tool_result block with matching tool_use_id. Then Claude continues, using the result. This full round-trip is one 'tool call cycle.'

Here's a full weather-tool exchange, end to end.

# 1. First call — Claude decides to use the tool
r = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    tools=[weather_tool],
    messages=[{"role":"user","content":"Weather in Toronto?"}]
)
# r.stop_reason == "tool_use"
# r.content[0] is a tool_use block

# 2. Execute the tool ourselves
tool_use = r.content[0]
result = get_weather(tool_use.input["city"])  # our function

# 3. Send the result back
r2 = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    tools=[weather_tool],
    messages=[
        {"role":"user","content":"Weather in Toronto?"},
        {"role":"assistant","content":r.content},
        {"role":"user","content":[{
            "type":"tool_result",
            "tool_use_id":tool_use.id,
            "content":str(result)
        }]}
    ]
)
# r2 is Claude's final answer using the result

The model doesn't 'run' anything. It outputs structured text that says 'I want to call this tool with these arguments.' You are 100% in control of whether and how that tool actually executes. Tools are a way for the model to REQUEST an action, not to perform one directly.

One tool call cycle isn't yet an agent. But loop it — keep sending tool results back and letting the model decide the next action — and you have one. Every agentic system on the exam is built on top of this basic tool_use / tool_result mechanic.