What is Guardrails?

Constraints that keep an agent's actions within safe bounds — what tools it may call, what it may spend, and what requires human approval.

How it works

An agent with real tools can take real actions, so you bound it: allow-list tools, cap spending, require confirmation for irreversible steps. Guardrails are what make it safe to let an agent run without watching every move.

Example

In practiceA purchasing agent can browse and compare but must get approval before any transaction over a set amount.

What should always require approval?

Anything irreversible or costly: spending money, sending messages publicly, deleting data, or acting on another system.

Go deeper: Deploy Your First AI Agent

The complete, step-by-step guide — every command, config, and the reasoning behind it.

Get the guide