Guardrails enforce ethical, legal, and operational limits on what an agent can access or do. They combine policy-as-code, rate limits, and content filters to prevent data leaks, bias, or unauthorized actions — the behavioural complement to identity and authorization systems.