An AI agent that can act on its own can also act wrong on its own. Guardrails are how you give it room to work without giving it room to do damage.
Why it matters
The value of an agent is that it acts without a human clicking approve every time. Guardrails make that safe. They set the outer edge, so a mistake stays small instead of turning into a breach.
Where teams get it wrong
Many teams treat a system prompt as a guardrail. A prompt is a request, not a limit. Real guardrails sit outside the model, in code the agent can't talk its way past. If a user or a poisoned page can argue the agent out of the rule, it was never a guardrail.
One example
Give a support agent a hard cap on refund size. Ask it nicely for more and it should still be blocked, because the cap lives in the payment system, not in the prompt. That's a guardrail doing its job.