Wire the real model — swap fake_llm for the Anthropic SDK shape (step 9/9) · capstone

Checkpoint

One last thing before we move on. Same surface as a write step — but the lesson doesn't complete until this passes.

Final drill. Build the full agent loop using the real-SDK shape. Write run_agent(question, max_iters) that:

Loads the API key via os.getenv("ANTHROPIC_API_KEY") (mocked to a known value for the demo). If missing, return {"ok": False, "error": "missing key"} immediately.
Otherwise, builds messages = [{"role": "user", "content": question}] and loops up to max_iters times calling fake_create(messages) (which returns a MockMessage).
On response.stop_reason == "end_turn": collect all text blocks from response.content, join them with " ", return {"ok": True, "answer": <joined text>, "iters": <iteration count>, "tokens": <input + output>}.
On response.stop_reason == "tool_use": for each tool_use block, dispatch through TOOLS[block.name](**block.input) to get a result string. Append the assistant turn (passing response.content directly through) AND a user turn with tool_result blocks. Continue.
On cap exhaustion: return {"ok": False, "error": "capped", "iters": max_iters}.

Token tracking: sum response.usage.input_tokens + response.usage.output_tokens across every call. Include the running total in the success return.

Expected output:

ok=True iters=2 tokens=120 answer=Found Tokyo's best ramen, 2026 guide.

⌘↵ runs the editor.read, then continue.

Checkpoint

One last thing before we move on. Same surface as a write step — but the lesson doesn't complete until this passes.

Final drill. Build the full agent loop using the real-SDK shape. Write run_agent(question, max_iters) that:

Loads the API key via os.getenv("ANTHROPIC_API_KEY") (mocked to a known value for the demo). If missing, return {"ok": False, "error": "missing key"} immediately.
Otherwise, builds messages = [{"role": "user", "content": question}] and loops up to max_iters times calling fake_create(messages) (which returns a MockMessage).
On response.stop_reason == "end_turn": collect all text blocks from response.content, join them with " ", return {"ok": True, "answer": <joined text>, "iters": <iteration count>, "tokens": <input + output>}.
On response.stop_reason == "tool_use": for each tool_use block, dispatch through TOOLS[block.name](**block.input) to get a result string. Append the assistant turn (passing response.content directly through) AND a user turn with tool_result blocks. Continue.
On cap exhaustion: return {"ok": False, "error": "capped", "iters": max_iters}.

Token tracking: sum response.usage.input_tokens + response.usage.output_tokens across every call. Include the running total in the success return.

Expected output:

ok=True iters=2 tokens=120 answer=Found Tokyo's best ramen, 2026 guide.

this step needs the editor

on desktop today; in the app (coming soon). save your spot and we'll bring you back here when you're ready.

open this same url on a laptop to keep going today.