Production agents need multiple layers of safety. Input validation catches malicious prompts. Output filtering prevents harmful responses. Rate limiting controls cost and abuse. Content moderation ensures appropriate responses. Human-in-the-loop adds oversight for critical decisions. Defense in depth is essential - no single guardrail is sufficient. Building safety from the start is easier than adding it later.