Guardrails are the measures that make an AI agent's behavior predictable and safe. Instructions in the prompt tell the agent what it can and can't do, but a model can break them. That's why key guardrails are enforced at the system level: the agent simply can't do anything it doesn't have permission for.
Types of guardrails:
- Access: only the data and actions it needs, with minimal token permissions
- Limits: number of steps, run time, request rate
- Filters: checks on incoming requests and the agent's replies
- Approval: a person signs off on risky actions
- Logging: a record of the agent's actions for review
What Pachca offers:
- A team agent sees public channels and only the private chats it's been added to
- You can set scopes on an employee's token, like read-only access to messages
- Administrators decide who can use the API and create integrations
- Actions taken through the API go into the audit log on the Corporation plan
- You can revoke a token or disable a bot, and access is cut off immediately
Learn more in the Help center
Read the article →For developers
Docs on dev.pachca.com →