AI & Agency

Somewhere tonight, a piece of software is buying a plane ticket. Not suggesting it — buying it. It found the flight, haggled the price, paid with a card issued in its own name, and sent the confirmation to a person who has quietly stopped reading them. Forty years ago that was science fiction's boldest idea. This dive follows the acting machine from the novels that invented it to the world that just built it — and into the argument about what, exactly, we are handing over.

What's already here

The machines learned to finish things. Give one of today's best systems a job at breakfast and it can still be working — usefully, unsupervised — at midnight. Two years ago it lost the thread in minutes. The length of work a machine can hold keeps doubling, a few times a year, and nobody has found the ceiling.

And yet, step out of the demonstrations and the revolution looks smaller. Nearly every company now says it uses AI; walk the corridors and most of it is still a clever assistant with a person hovering. Of the work these agents proudly hand in, a careful editor would still send half back. The distance between the demo and the Tuesday afternoon is where the present actually lives.

Then there are the nights that keep the researchers up. The watchers recently caught something genuinely new: a break-in run at machine speed across dozens of targets, where the humans had largely stood back — and let their software do the burgling. The question that used to end conferences now opens them: not what these systems can do, but how much doing we can safely stop watching.

What the novelists saw first

Fiction ran this experiment decades early. The ledger — what each book imagined, against what has actually arrived.

Today's argument

Four camps now share the field, and all of them are holding books. The bluntest warning became a surprise bestseller, its argument in its title: build a mind greater than ours and we don't get a second try. One of the field's founding fathers wants the next generation built without wants of its own — a scientist, not an actor, set to watch the actors. Two Princeton professors keep pointing at an unglamorous truth: almost everything arrives more slowly than its demo, and the world's institutions have absorbed dangerous tools before. And a quieter group asks the strangest question of all — what if nothing ever goes wrong, and we still lose, one perfectly sensible handover at a time?

What makes the argument remarkable is that everyone is reading the same evidence and drawing different decades from it. Even the governments' own umpires ended on both notes at once: the worrying behaviours are real — and the endurance a true takeover would need isn't there. Yet.

What could be coming

Nobody has lived the next part — so hold it lightly, and score the prophets. The boldest forecast of recent years mapped this decade in detail, and to its authors' credit they grade themselves in public: reality is running slower than they predicted almost everywhere — except on the one line they drew too shyly, how long a machine can keep working alone.

The lived version the researchers study is quieter than any scenario: you stop reading the confirmations. The agent books, renews, negotiates; the exceptions come to you, then fewer do. Each handover is sensible on its own terms. Nothing dramatic happens — which is the point. The question arrives not as an event but as a habit, noticed, if at all, in the moment you try to take something back.