The Hard Part of AI Agents Isn't Autonomy, It's Trust
A capable model can now read a support ticket, query three systems, draft a refund, and issue it before a person finishes their coffee. Getting an agent to act is no longer the interesting problem. The interesting probl…