What is an Agent Guardrails?

A guardrail is distinct from prompting the model to "please be careful" — instructions in a prompt are guidance the model can be talked out of, especially by adversarial input, while a guardrail is enforced in code outside the model's control. Examples include hard limits on which tools an agent can call, rate limits on how many actions it can take per minute, content filters that block certain categories of output, and mandatory approval gates on high-risk actions that no prompt can bypass. Good guardrail design assumes the model will eventually be wrong, tricked, or confused — through a bad instruction, a hallucination, or a prompt-injection attempt buried in a document the agent reads — and builds the safety net for that case rather than trusting the model to always self-correct. Guardrails should also fail closed: if a guardrail check errors out or can't complete, the safe default is to block the action, not to let it through.