Guardrails are the limits imposed on an agent: allowed tools, data scope, number of steps, budget, input and output filtering, action approval. They are enforced by the application's code, not merely requested in the prompt.
Example
Refuse on the server side any write action outside the user's project, whatever the model requests.