Setup

Three tools, then a checkpoint. If you already have any of them, skip ahead.

1. Install uv

uv manages Python versions, project environments, and packages — one tool instead of three.

brew install uv

(Linux: curl -LsSf https://astral.sh/uv/install.sh | sh)

Verify it:

uv --version

Any version number is a pass. Get in the habit: install, then verify. It turns silent failures into obvious ones.

2. Create the project

mkdir -p ~/agent-course && cd ~/agent-course
uv init --python 3.12
uv add ollama

uv init creates a project with its own isolated environment — your system Python is never touched. uv add ollama installs the one package this course needs: the client library for talking to a local model server.

You’ll run every file in this course with uv run <file> — no virtual-environment activation, no pip, nothing to remember.

3. Install Ollama and pull the model

Ollama runs open-source language models on your own machine. No account, no API key, no per-call cost.

brew install ollama
brew services start ollama
ollama pull qwen2.5:3b

The pull downloads about 2GB — a few minutes. qwen2.5:3b is a small model by today’s standards, and that’s deliberate: it’s free, it runs on a laptop, and its mistakes are visible, which is exactly what you want while learning how the machinery works.

Why not llama3.2?

It looks like the obvious choice and it does not work for this course. llama3.2 emits tool calls correctly but ignores the results — hand it data it asked for and it replies “I don’t have access to real-time data.” Tested head-to-head on the same task: llama3.2 scored 0/3, qwen2.5:3b scored 3/3. There’s a full write-up in When it breaks.

The lesson generalises: “supports tool calling” is not a yes/no property, and vendor capability tables won’t tell you this. Only a test does.

4. Checkpoint

ollama run qwen2.5:3b "Explain what a while loop does in one sentence."

If a sentence streams back, everything works. That text was generated on your machine — nothing left it, nothing was billed.

One thing worth noticing before you move on: the answer arrives word by word. That’s not a UI effect. The model genuinely produces one token at a time, each conditioned on everything before it. A language model is a machine that predicts what text comes next — everything in this course is built on top of that single primitive.

Step 1 — prove the model can’t know your data →