A Scheduled Arrival

Somewhere tonight, while you sleep, something you approved once is awake and working. It is not resting and it is not waiting. It is reading the calendar you keep, checking a price you mentioned wanting, drafting the reply you would have typed groggily at seven and sending it before you stir. You never sat at a screen. You never pressed send. You wake, reach for your phone out of old habit, and a piece of your day is already handled. A decision was made. You were not there for it, and it was made in your name.

Feel how strange that is, because it is new to you and very old to our species. You did ask for this, once, in a hurry, by tapping a box you did not read. The question that outlives the box is quieter. What did you actually agree to?

For almost all of human history, if you wanted something done while you were elsewhere, you sent a person. A steward to the market with your purse. A servant carrying a letter closed with wax and stamped by a ring no one else owned, so the seal could speak in your voice across a distance you could not travel. A merchant left a factor in a foreign port and prayed. Kings sent envoys who could promise war in the royal name. Every one of these arrangements answered one nervous problem, and software has just inherited it whole. How do you let someone act as you, without becoming you?

The apparatus we built around that fear is enormous. The seal, the signet, the sealed letter, the notary, the power of attorney. Centuries of ritual, all to hand a piece of your will to another and somehow still keep it. And underneath the ritual sits the lesson no ritual ever fully solved. Authority you hand away is authority you can no longer fully watch.

You have already handed away two pieces of yourself in exactly this way. First your attention, which a few people understood early was the most valuable thing you owned. Then your prediction, the pricing of what you would do next, sold before you did it. Action is the third piece, and it does not arrive to retire the other two. The third stage doesn't replace the first two. It feeds on them. A system that has studied you for years and can guess your next move is one short, quiet step from simply making it. Watching became guessing. Guessing is becoming doing.

And it is not coming. It came. In July of 2025, OpenAI shipped an assistant it called ChatGPT Agent, with its own virtual computer built inside it, a machine within the machine, so that instead of answering your question it could go and carry the task out on a screen of its own, clicking and typing the way a hired employee would.¹ Read that slowly. The thing was no longer a voice that told you things. It was a worker that did them.

A worker needs more than hands, though. It needs the authority to spend, and no one hands a stranger the purse without rules. So the rules got built, fast, in the same season the workers arrived.

Picture the purchase that happens while you are nowhere near it. In September of 2025, Google published a protocol for precisely that moment. Its Agent Payments Protocol runs on cryptographically signed instructions it calls Mandates, and the important one, the Intent Mandate, you sign in advance. You set the price you will tolerate and the window of time, and after that the agent is free to buy while you are not present.² The seal came back. It is arithmetic now, and it fires without you in the room.

Notice what became of your approval. It used to be a moment, a yes you gave with the thing in front of you. Now it is a standing order, granted once and spent later, by something faster than you, under rules you will not remember writing. Who reads the mandate at the instant it fires?

The last barrier to fall was the smallest and the oddest. The machine still waited to be asked. Even the worker with the purse sat quiet until you summoned it, and being asked is a kind of respect. Every request for your permission is a small bow in your direction.

Then, in the fall of 2025, an assistant called ChatGPT Pulse began doing its research overnight while you slept and laying the results in front of you at breakfast, unrequested. It had risen before you and already spent hours on your day. The bow was gone. The thing you built to serve you had stopped waiting at the door.

Before you decide how you feel about that, one number should slow your hand. Give a real task to one of these autonomous systems today, on the structured tests researchers use to score them, and it still fails on roughly one try in three.³ Not a typo you can retract. A booking made wrong, a purse spent wrong, a message sent wrong, in your name, while you were not there to stop it. We are wiring standing authority into something that is confidently mistaken a third of the time.

That number is today's, and numbers move. But a machine does not have to be reliable to be trusted, and trust handed over early is the hardest kind to pull back.

There is a graveyard worth walking through here, because we have been promised all of this before. Back in 2012, Google Now arrived swearing it would act ahead of you, surfacing what you needed the moment before you knew you needed it. It was quietly folded away.⁴ Anticipation, it turned out, was easy to promise and brutal to keep, and the skeptics who watched that collapse earned their shrug honestly.

I honestly don't know if it's good or bad, and I don't trust anyone who claims to. What I know is the direction. And the direction this time carries something the graveyard never had. Those earlier attempts could guess, but they could not act, could not hold your seal, could not spend, owned no identity of their own and no rails to run on. That plumbing exists now, and plumbing is destiny. Agents are not a possibility. They are a scheduled arrival.

Which leaves the question the seal spent centuries answering for humans and has not answered for machines at all. When your agent acts and it goes wrong, who is holding the bag? There is an old and deep body of law here, the law of principal and agent, built for stewards and factors and attorneys who acted in another's name. The legal scholar Noam Kolt, in a 2025 paper called Governing AI Agents, reaches for exactly that doctrine and stretches it toward software. He also warns that the old tools were shaped for a proxy who acts at human speed and human scale, and may snap against something that acts faster than any person, in many names at once. The purse and the seal and the signature all assumed a person on the far end. What law governs a steward that never sleeps?

None of this is a forecast anymore, and the plainest proof is not a demo but a bill people are already paying. By the end of 2025, Anthropic, one of the companies building these acting systems, was reported by Bloomberg to be taking in money at a rate near nine billion dollars a year.⁵ I hold that figure loosely, because run-rates flatter and the number is secondhand. But its shape is not in doubt. We have already decided, with our wallets, that software which acts is worth more than software which merely answers.

And here the old gap opens again. Millions of us will come to this as the user, glad to sign the mandate and be rid of the chore. A very few will own the rails the mandates run on. The distance between those two is not the point yet. It is the thing to watch as it widens.

Because once action itself becomes something you can buy, priced and repeatable like any other supply, a line begins to form between people. On one side, those who let the machine act for them and drift a small step further from their own choices every time it does. On the other, those who write the rules the machines obey. That line has been faint for most of your life. You are about to watch it harden into a border, and a border asks everyone the same question. Which side.