The Helpful Agent Wants Your Inbox, Your Calendar and Your Bank Account

Every lab is now selling you the same creature: an agent that never sleeps, never forgets, and — if you let it — does your shopping. The pitch is a colleague who works while you rest. The price is everything you own.

That's not a bug in the design. It is the design.

Meta, OpenAI, Anthropic and xAI are all racing to put a consumer agent in your life, and the deal is identical in every case: the agent gets better the more it knows about you, and the tasks that actually matter — sorting email, booking flights, buying things — demand the keys to your inbox, your calendar, your ID and your bank account. The more useful it is, the more of yourself it holds. Nobody in the race is pretending otherwise.

The committee, not the servant

The marketing job titles are revealing. Claude Duet, OpenAI's Dots and xAI's Grok Bot want to be your colleague or chief of staff. Meta's Muse and Spear Street's Instinct pitch themselves as personal assistants. But the AP's reporting is blunt: none of them is incapable of working outside its marketed lane. The labels are positioning, not capability — and the real answer from people who actually use these things daily is to run a stack of agents from different labs, each checking the others' work.

"One agent does the work and then I let the other one critique and criticize it," said Clara Gigling, founder of Oudae Labs. "That way, they find mistakes, correct themselves and become better."

So the future of personal computing isn't one brilliant servant. It's a committee. And committees need managing — by you, the only member who can't be replaced.

The fine print is the product

The safeguards are real, and some are genuinely thoughtful. Anthropic's Duet runs on its own cloud computer, must show you its intended step and get approval before sending a message, making a payment or deleting anything, never sees your passwords, and pays with a one-time card so your real number stays hidden. Trading securities and moving money between accounts are off-limits even with approval.

Meta's Muse has its own dedicated cloud computer too — and its own privacy policy says the setting that lets your interactions train its AI models is on by default. You have to switch it off. Deleted items may still be "remembered"; only a full Reset wipes everything. The fix Meta announced — a Confidential VM encrypted with a key only you hold, "so not even Meta can access it" — was still described as coming in the future as of early November.

Read that twice: the privacy-preserving version of the product is a promise, not a feature.

Start with one annoying thing

The honest advice from people using these agents is refreshingly unglamorous: start with one annoying, repeated task you've been putting off — a long email, a flight you keep not booking. Be specific. Give the agent a detailed goal, the context, and the tools. Then try the others and see what fits.

That's also the quiet confession at the heart of this market: nobody can tell you which agent fits, because the fit depends on the work — and on the bargain. The useful agents are the ones that ask for the keys. Approval gates, one-time cards and memory controls make that bargain safer. But the bargain is still yours to make, one agent at a time.

The real question was never which agent is smartest. It's how much of yourself you're willing to hand over — and to how many of them.