Definition
Guardrails are safety mechanisms that constrain AI agent behavior to prevent harmful, unintended, or out-of-scope actions. They include input validation (filtering harmful prompts), output filtering (blocking sensitive data), action limits (restricting which tools an agent can use), and human-in-the-loop checkpoints for high-risk decisions.
Example
An AI agent that can draft emails but cannot send them to anyone outside the company without manager approval. Or an agent that can read financial data but cannot modify records without authorization.
Why it matters
Guardrails are what make AI agents safe for business use. Without them, agents might take actions that violate policy, leak data, or cause financial harm.