The widespread design assumption that users must see what the agent is doing is a legacy of treating agents as untrustworthy; once an agent can hook into data and just works, close supervision becomes unnecessary and undesirable—users want it to be fast, trustworthy, and dependable.