👆 Tap a line → plain English
🔎 go deeper → extra detail
# dim italics = a comment
Guardrail
block bad input before tools
Eval
pass rate on real cases
Run it right here. Every cell runs in your browser. Run them in order, 1 → 4; they share memory. First Run downloads Python once (~10–15 MB).
This is the plumbing that makes an agent trustworthy. The guardrail, trace, and eval here are real Python you'd keep almost unchanged in production — only the agent's “answer” line would become a real model call. Trust isn't a model feature; it's this code around the model.