Guardrails

Guardrails are the rules and filters placed around an AI agent that constrain its inputs and outputs to keep its behavior safe, on-policy, and within scope.

Guardrails can block disallowed content, validate that outputs meet a schema, restrict which tools an agent may call, and catch prompt-injection attempts before they reach the model.

They are part of agent governance and a factor in whether an agent is safe to deploy, not just whether it is accurate.