Definition
A guardrail is any control that helps an agent stay within acceptable boundaries. It can be a prompt rule, policy retrieval, schema validator, approval workflow, or blocklist for prohibited claims.
AI Agent term
Guardrail
A rule, validation step, policy, or review gate that limits risky agent behavior.
Why it matters
Model choice alone does not make an agent safe. Guardrails reduce the chance that a fluent answer becomes a harmful business action.
Example
A refund guardrail blocks the agent from promising compensation unless an approved policy condition is met.