Same loop, real model

Lesson 1 wrote the agent loop on top of fake_llm — a function that returned canned response dicts. The whole point of building it that way was that the loop is the load-bearing part. The model is swappable.

Here's what the swap actually looks like. Two lines change:

import os
from anthropic import Anthropic

client = Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))

That's it. The messages.create(...) call returns a response object whose .stop_reason, .content, .content[i].type, and .content[i].text access patterns are identical to the dict-shaped mock you built. The loop you wrote in lesson 1 runs on real production traffic with no other code changes.

Three pieces this lesson locks in:

1. The key lives in the environment

os.getenv("ANTHROPIC_API_KEY") — never hardcoded, never in source, never committed. Chapter 18 set this up; this is the chapter where it actually matters because the request goes out to a real billed endpoint.

The two reflexes you'll fight:

• Putting the key in the source for "just this one test." It ends up in git history. You rotate the key. Two days later you notice the bill. By then it's too late.
• Putting the key in os.environ["ANTHROPIC_API_KEY"] (bracket access). This raises KeyError if the env var is missing. os.getenv returns None and lets you give a clear error upstream. Always getenv, never bracket.

2. The SDK response is OBJECTS, not dicts

The mock you wrote in lesson 1 used dicts (response["stop_reason"], response["content"][0]["text"]). The real SDK uses object access (response.stop_reason, response.content[0].text). Almost identical, slightly different.

Every block in response.content has a .type field. For text blocks, .text. For tool_use blocks, .name, .input, .id. You must check .type before reading the type-specific fields, or you'll get AttributeError on the first tool-using turn.

3. The loop doesn't care

The genuinely good news: the loop you wrote in lesson 1 — the while, the stop_reason branch, the tool dispatch, the messages.append(...) — all runs unchanged. The mock you'll build in step 8 has exactly the real SDK's attribute shape. Lift the loop, change the call site, ship it.

What you'll build

A MockResponse with MockTextBlock and MockToolUseBlock classes whose attribute access matches real anthropic.types.Message. Then a fix for the most common KeyError bug (bracket access on env vars). Then a fix for the AttributeError bug (reading .text on a tool_use block). Then a real shape adapter that takes either a real SDK response or a mock and runs the loop unchanged.