Exercise 1: One Tool, One Complete Round Trip¶

Corresponds to Exercise 1 in Stage 3 — Tool Use & Your First Agent Loop.

This exercise does one thing: the model says “call get_weather,” your Python program validates the arguments, executes the tool, and returns the result to the model. After running it, you will see:

Question → Tool Call → Program validates and executes → Tool Result → Final answer

Tool Call is a tool request from the model. Tool Result is the result returned by your program after execution. A model request does not mean it has permission to execute code directly.

First action¶

In PowerShell, copy and run:

ollama pull qwen2.5:3b

Path A: Ollama (local, API cost `$0`)¶

cd examples/stage-3/01-function-calling
python -m pip install -r requirements.txt
ollama serve
python starter.py

If ollama serve says the port is already in use, Ollama is usually already running; leave that window open and run python starter.py in another PowerShell window.

This path uses the OpenAI Python SDK connected to http://localhost:11434/v1; data is not sent to the OpenAI cloud.

Path B: Anthropic (requires an API key)¶

cd examples/stage-3/01-function-calling
python -m pip install -r requirements.txt
$env:ANTHROPIC_API_KEY = "your-key"
python starter_anthropic.py

The program uses the pinned model ID claude-haiku-4-5-20251001, so the teaching result does not silently change when a model alias moves.

Budget reminder: reserve a $0.05 cap for each real run. Actual cost depends on token count:

input tokens × $1 / 1,000,000 + output tokens × $5 / 1,000,000

Tool Use also adds system-prompt tokens; do not present decimals based on a no-token assumption as guaranteed prices. Price checked on 2026-08-27.

macOS/Linux commands

cd examples/stage-3/01-function-calling
python -m pip install -r requirements.txt
export ANTHROPIC_API_KEY="your-key"
python starter_anthropic.py

Free self-check¶

These tests use fake model responses: they do not connect to Ollama or call the Anthropic API.

python test.py
python test_anthropic.py

You should see all pass twice. The tests also deliberately send bad JSON, extra fields, and an unknown tool to confirm the program blocks them first.

What you are protecting¶

Allowlist: only get_weather can execute; a model-generated different tool name is rejected.
Argument validation: city cannot be empty, unit must be celsius, and extra fields are rejected.
Result matching: every result carries the original tool_call_id or tool_use_id.
Error marker: the Anthropic path adds is_error: true on failure so the model knows it is not a normal result.

Completion conditions¶

Path A or Path B succeeds at least once.
Both offline tests show all pass.
I can explain in my own words: “The model only makes a request; the program actually executes it.”
I can point to where the program validates the tool name and arguments.

Official references¶

Docs and SDKs checked on 2026-08-27.

📚 Want the chapter-length version? This folder teaches only the smallest first loop. Continue with: - datawhalechina/hello-agents: a chapter-based Chinese Agent course; use this exercise as the tool-calling starting point. - Anthropic Tool Use Cookbook: official notebooks that grow from one tool to multiple tools. - Stage 3 Curated Projects: return to the learning map and choose the next resource.