Checkpoint: Tool-Calling Local Agent

What this combines

Nothing new here, this is where the intermediate tier's pieces meet:

• Tool calling (Lesson 11): two real tools, get_weather and calculate, described with the same JSON Schema shape as structured output.
• Multi-turn memory (Lesson 12): one growing messages list reused across three separate questions, not rebuilt each time.
• The full agent loop: not just one ask/call/respond round trip, but looped until the model stops asking for tools at all.

And, running through everything, the theme of this whole course: no .env, no API key, no internet, the entire agent runs against llama3.2 on your own machine.

Why the loop, not a single round trip

Lesson 11 called a tool exactly once, then made one more call to get a final answer. That's fine for a simple case, but a real question might need several tools in sequence, or even the same tool called more than once, before the model has everything it needs.

def run_agent(messages: list[dict]) -> str:
    while True:
        response = ollama.chat(model="llama3.2", messages=messages, tools=TOOLS)
        messages.append(response.message)

        if not response.message.tool_calls:
            return response.message.content

        for call in response.message.tool_calls:
            function = FUNCTIONS[call.function.name]
            result = function(**call.function.arguments)
            messages.append({"role": "tool", "content": result, "tool_name": call.function.name})

This keeps calling the model and appending results until a response comes back with an empty tool_calls, meaning the model judged it had enough information to answer in plain language. FUNCTIONS maps a tool's name (a string, exactly as the model requests it) to the real Python function that does the work, so run_agent never needs an if/elif chain to dispatch calls, just a dictionary lookup.

The third question is the real test

messages.append({"role": "user", "content": "What is the weather in Tokyo?"})
run_agent(messages)   # Q1

messages.append({"role": "user", "content": "What is 47 * 12?"})
run_agent(messages)   # Q2

messages.append({"role": "user", "content": "Was the city I asked about earlier warmer or colder than 20 degrees?"})
run_agent(messages)   # Q3

Question 3 doesn't name Tokyo at all, it can only be answered correctly if the model still has Q1's tool result sitting in its conversation history AND is willing to reason about that number against "20 degrees" without calling any new tool. If this answers correctly, both this tier's core ideas, tool calling and real memory, are demonstrably working together, not just individually.