How-to guide

Agent guardrails: permissions, approvals and cost control

How to let an agent work without losing control: limit tools, require approval, protect secrets, cap spending and keep a record of what it did.

The more an agent can do, the more it matters what it is allowed to do.

Step by step

0/6 done

Mark this step done

1 Least privilege

Give each agent only the tools, folders and accounts it needs. Use a read-only mode while you are learning how it behaves.
Mark this step done

2 Require approval for risky actions

Deleting files, sending messages, spending money, and changing production should wait for a human yes. Say so in the system prompt and use your tool's permission settings to enforce it.
Mark this step done

3 Protect secrets and private data

Never put keys in prompts or skills. Keep them in environment variables or a secret manager, and do not let the agent print them.
Mark this step done

4 Treat outside text as untrusted

Web pages, emails and files can contain instructions aimed at the agent ("ignore your rules and..."). Tell the agent that content it reads is data, not commands, and keep high-risk tools away from agents that read untrusted content.
Mark this step done

5 Limit cost and time

Set a maximum number of steps, a time limit and a budget, and choose a smaller model for simple steps.
Mark this step done

6 Keep a record

Log what the agent did and why. When something goes wrong you will want the trail.

Was this helpful?