What is a guardrails?
Guardrails are the rules and limits around an AI agent that keep it inside safe bounds, such as how much it can send, what it must refuse and when a person must step in.
In plain words
A guardrail is like the barrier on a mountain road. It does not slow a careful driver down, but it stops a slip from becoming a disaster. For AI, guardrails keep a helpful agent from doing something costly, rude or risky.
Some guardrails are rules you write, such as "never promise a refund" or "hand over anything about health". Others are built into the tools the agent uses, so they hold even if the agent is confused or is asked to break them.
The second kind is the stronger one. An instruction can be misread or forgotten. A limit enforced by the service itself cannot be talked around, because the service simply refuses.
Good guardrails explain themselves. When an agent is stopped, it should learn why and when it can try again, so it can tell you rather than quietly failing. You do not need many guardrails to start.
A handful of clear ones, such as a ceiling on how much can go out in a day and a rule to hand over anything sensitive, already cover most of what can go wrong for a small business.
How it works in Spaces
Spaces enforces guardrails on its own side: a gap between messages (WhatsApp 10 seconds, Instagram 15 seconds, LinkedIn 20 seconds, Messenger 20 seconds, X 20 seconds), daily limits per account, a four-week warm-up for new accounts, and some messages that are always refused, such as the same cold message to many people or anything to someone who said stop.
On top of that, you can hold drafts for approval, hand a chat to a person or pause all sending.
Related terms
- Pacing: Message pacing means leaving a gap between messages and capping how many go out in an hour and a day, so an account sends at a speed a person could.
- Daily limit: A daily limit is the most messages, or new conversations, an account is allowed to send in one day before sending stops until the next day.
- Warm-up: Warm-up means starting a new or newly connected account with only a few new conversations a day and increasing slowly over a few weeks, so the platform sees normal behaviour rather than a sudden burst.
- Human in the loop: Human in the loop means a person stays part of the process, checking, approving or taking over, instead of leaving the AI to act entirely on its own.
FAQ
Can the agent get around the guardrails?
No. Spaces applies its limits on its own side and refuses what breaks them.
Do guardrails make the agent less useful?
For everyday replies you will rarely notice them. They matter on bulk jobs and risky requests.
Can I add my own rules?
Yes. Write them in your agent instructions, such as when to hand a chat to you.
Do guardrails remove all risk?
No. They lower it. Each platform still makes its own decisions about accounts.
Useful next: Account safety, Sending limits, Docs. Every term: the glossary.