You just taught the model to return what you expect. Now defend against when it does not. Guardrails are the checks between the model's output and the rest of your system: validate the response matches your schema, reject or retry on malformed or off-spec output, and never let an unchecked model response flow straight into something that acts on it.

This is the same muscle as structured outputs, from the other side. Introduced here as a day-one fundamental; the agent-specific version of the threat - prompt injection through tool use - comes back in Week 3 where it gets genuinely dangerous.