Guardrails are how an agent stops being a liability. They cover the things the agent must not do: discuss medical advice, quote unapproved prices, fire a refund above a threshold, talk about a competitor, hand out a phone number.
Useful guardrails are not a wall of natural-language policy. They are typed checks at the tool boundary (the refund tool only accepts amounts under $200), at the conversation boundary (certain topics route immediately to a human), and at the audit boundary (every action writes a row that a compliance officer can read).
They are also testable. A serious vendor ships a regression suite that drives the agent at every guardrail and verifies the right thing happens. Without that, "we have guardrails" is a marketing claim.
Vorel guardrails are typed at the tool layer and verified against a per-customer golden set before any change ships.