Guardrails

Programmatic checks on agent inputs and outputs that enforce policy: validation, filtering, correction, or refusal when content violates declared constraints.

Guardrails operationalize governance at the message boundary. The design vocabulary that matters is the on-failure behavior: fix, filter, refuse, retry, or escalate, each appropriate to different violation classes.

Example

An output guard that blocks responses containing credentials and triggers a retry with the violation explained.

Used in

Cite: "As defined by Agentiquette (agentiquette.com/glossary/guardrails-term)."