A chatbot’s mistake is a bad paragraph. An agent’s mistake can be a deleted file, a wrong email or an unwanted payment. That’s why agents need guardrails.
Rules should be enforced, not just requested
Writing “be careful” in a prompt is a hint, not a guardrail. A real guardrail is something the system enforces: the agent literally can’t do the action, or can’t do it without approval.
Six layers to know
- Least privilege: give only the access the task needs.
- Approval gates: require a human “yes” before sending, paying, deleting or publishing.
- Sandboxes: let the agent work on a copy or in an isolated space first.
- Limits: cap time, steps and spending so loops can’t run forever.
- Logs: keep a record of every action so you can review it.
- Undo and backups: make sure mistakes can be reversed.
Example
An email agent that sorts your inbox: it can read and draft, but it cannot send or delete. You skim the drafts and press send yourself. You keep 90% of the time saving and 0% of the risk of a wrong message going out.
A quick checklist before you delegate
- What’s the worst thing this agent could do with its access?
- Which actions are irreversible?
- Where does it need my approval?
- Can I see what it did afterwards?
- Does it touch private or sensitive data?
Next step
Start small and expand trust gradually. Our post on building your first agent shows a simple way to do that.