How this course works

Real shapes, instant runs

Real models need the internet and an API key, and they answer a little differently every time. That's bad for learning: you want to run your code a hundred times and see exactly what changed.

So in the browser you'll use a stand-in model: a small Python class with exactly the shapes of the real Claude API. client.messages.create(...) returns a response with .content blocks and a .stop_reason, just like the real SDK.

In module 12 you swap in the real client on your computer. Because the shapes are the same, your agent keeps working.

The stand-in's code runs first in every program, folded away under Runs first so you see only what matters. Open it any time: it's 30 short lines.

One call to the stand-in

def brain(messages):
    return Response(text("Hello! You said: " + messages[-1]["content"]))

client = Client(brain)
reply = client.messages.create(model="stand-in", max_tokens=200, messages=[{"role": "user", "content": "hi"}])
print(reply.stop_reason)
print(reply.content[0].text)

1. Line 1: A brain decides what the model says. This one just repeats your message back.
2. Line 4: The client wraps the brain, like `anthropic.Anthropic()` wraps the real model.
3. Line 5: One call: a list of messages in, a response out.
4. Line 6: No tool was used, so the model is done: `end_turn`.
5. Line 7: The reply's text lives in its first content block.

Done, or not done?

def brain(messages):
    return Response(text("Let me check."), tool_use("call_1", "search", query="weather Pune"))

reply = Client(brain).messages.create(model="stand-in", max_tokens=200, messages=[{"role": "user", "content": "Rain today?"}])
print(reply.stop_reason, len(reply.content), reply.content[1].name)

What does this print?

• tool_use 2 search
• end_turn 2 search
• tool_use 1 search