Control and Permissions

The more independent an agent is, the more important it is to decide in advance what he can do without asking, and what he can do only with your permission. Actions are divided into reversible (compose a draft, collect data, calculate) and irreversible (send an email, delete a file, pay, publish). The mechanics of control are simple: we resolve reversible things automatically, and set irreversible ones for confirmation. Insider technique “dry run”: let the agent first show what exactly he is going to do (to whom is the letter, what text, what amount), and wait for your “yes”. You catch a mistake before it happens in the real world. The second technique is the stopping rule: describe in advance in what cases the agent is obliged to stop and call you (the amount is above the limit, there is not enough data, the action affects the client). A typical mistake is to give the agent complete autonomy on risky operations for the sake of “speed”: one wrong step with sending or deleting cannot be rolled back. Autonomy is increased gradually: first, the agent makes a suggestion, you confirm, and auto-actions are expanded only where the cost of an error is small.