With a normal feature, a mistake is a wrong number on a screen. With an agent, it’s an email already sent or a payment reminder already fired off. Picture an invoicing tool that automatically sends a firm reminder to a customer who’s been loyal for years. The agent can be right about the data and still wrong about the relationship.

I’d judge an agent less by how smart it is and more by how cheaply I can catch and reverse its mistakes: a clear log of what it did, a short delay before it acts, a one-click undo. New hires get a probation period, and I think agents should earn autonomy the same way, starting with suggestions, then drafts, then acting alone.