Tool Calling

Where we left off

Lesson 13 built a calculator tool and called it directly, no AI involved. This lesson connects that same tool to a model, and walks through what actually happens when an AI uses a tool.

Connecting the tool to the model

model_with_tools = model.bind_tools([calculator])

bind_tools does not give the model the power to run Python. The model can never execute code, it only ever produces text. What bind_tools does is describe the tool (its name, description, and argument schema, exactly what Lesson 13 printed) to the model, so that when the model wants to use it, it can output a special structured message meaning "please run calculator with this input", instead of guessing an answer in plain words.

The three rounds of conversation

Talking to a model with tools isn't one message and one reply, it's a back-and-forth, usually three steps.

Round 1: ask the question.

ai_message = model_with_tools.invoke(messages)

We send: "What is 847293 multiplied by 3821?" The model recognizes it needs exact arithmetic, and instead of answering in words, it replies with a request: "call calculator with expression = '847293 3821'"*. This request lives in ai_message.tool_calls, a list, since the model could ask for more than one tool call at once.

Round 2: we run the tool ourselves, and report back.

for call in ai_message.tool_calls:
    result = calculator.invoke(call["args"])
    messages.append(ToolMessage(content=result, tool_call_id=call["id"]))

The model asked, but it cannot press the button itself, our program has to actually run calculator.invoke(...). This is the important idea to hold onto: the AI decides, but your code executes. The AI never directly touches your computer, your files, or the internet. It only produces a request; your program chooses whether to honor it, and is the one that actually does it.

ToolMessage wraps the result and gets added to the conversation. tool_call_id is like a receipt number, it tells the model "this result answers the specific request you made with this ID" (useful when there are multiple tool calls at once, so results don't get mixed up).

Round 3: let the model turn the number into a sentence.

final_response = model_with_tools.invoke(messages)

We send the entire conversation so far: the original question, the model's own tool request, and the tool's result. The model reads all of that and writes a normal, human-sounding final answer, using the exact number the calculator produced instead of guessing.

Why does messages keep growing?

Notice messages starts as one item and ends up with four: the question, the AI's tool request, our tool result, and the final answer. Each .invoke() call doesn't remember anything on its own, the model has no memory between calls. Every time we call .invoke(), we hand it the entire conversation so far, or it won't know what already happened. This will matter a lot in Lesson 17.