Agents can carry more of a life. What must they never quietly take over?
AI assistants are becoming agents. They do not merely answer questions. They remember across time, use tools, speak to other people, spend money, schedule work and act inside the systems through which a life is conducted.
For that to be useful, they must carry far more than a person could supervise step by step. A person who must approve every classification, calendar move or recurring payment has not gained an agent. They have acquired another job.
But broad autonomy creates a harder problem than privacy or permission. Who controls the context through which the machine understands the person? Who decides what currently matters? What may the machine say in the person's name? Which consequences can it create? How does the person revoke it, change direction or leave without losing the continuity the machine accumulated?
What must remain under human authority when agents carry the rest?
Sovereignty is not constant supervision.
The answer cannot be another approval screen. Clicking yes to every operation would make agency useless, while clicking yes to a boundary nobody understands would make consent fictional.
The human ratifies the mandate. The machine may settle ordinary instances inside it. The machine may not create the mandate, silently expand it, conceal which mandate authorized an action or cross a declared boundary without returning to the appropriate human authority.
Sovereignty here does not mean solitary control or maximal independence. It means machine power remains anchored in the person or people who legitimately hold the relevant authority, including when that authority is shared, professional or institutional.
This moves human authority away from performative clicking and toward five deeper controls: the evidential floor, the personal frame, the grant of authority, the consequential commit and the path of return.