Agent permissions
Letting AI agents act safely on a user's behalf
What it is
Agent permissions control what an AI agent can do when it acts for someone: which tools and APIs it can call, whose data it can touch, and which actions need a person to approve first. The safe pattern is to give the agent the acting user's permissions and no more, keep credentials out of the model's reach, log every action, and require approval before anything destructive or irreversible.
Buyers are comfortable with AI that drafts and suggests. They get nervous when it can send, delete or pay, and they'll ask exactly how that's contained.
What it takes
About 5–10 engineer-weeks to build in-house, or 2–4 using Arcade, Oso, Permit.io.
Answer these first
- Product: Which actions can an agent take on its own, and which need a person to approve?
- Engineering: Does the agent act with the user's permissions, or broader ones of its own?
- Engineering: Where are credentials for connected tools stored, and can the model ever see them?
- Security: How would we stop an agent that's been manipulated by prompt injection?
- Support: Can a customer see a log of every action an agent took?
- Legal: Who is responsible when an agent takes the wrong action?
This page works best with JavaScript on. Every answer also has its own address, like /what/scim.