An agent is a model with hands.

On the call, we would ask

• Which of your systems would it need to reach into, and who controls that access?
• What is the worst thing it could do if it misread a step?
• Who approves an action today, and would they still approve it?
• How would you know afterwards exactly what it did?