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)."