The hard part of managing AI turned out to be AI's jobThe Dot and the Swarm

The author expected that organising AI agents would take years of careful human design; the models turned out to be good at it, because most management exists to fix problems that only people have.

The author has been writing about AI for years and thought his predictions had held up. Recently he admitted he got one big thing wrong.

He had assumed that directing a group of agents would work like running a team: who does what, who reports to whom, all specified by humans, with years of trial and error to get it right.

The same trap, again

There is a rule in AI that keeps proving true: problems people are sure will need elaborate rules and careful design end up falling to brute force and better models.

People once built whole systems to feed AI the right information; models learned to go find it. People once wrote long templates walking AI through a task step by step; newer models plan the steps themselves.

Organising was next. Management has been studied for thousands of years without anyone cracking it, so it looked like something humans would have to design, at least for a while. It turned out to be one more thing AI can learn.

A thousand assistants minding your business

The top app in the store right now is Meta's Muse, a personal agent that promises to do work for you. OpenAI's dots competes with it, and a crowd of others do roughly the same.

They give an AI access to a computer and connect it to your email and financial accounts. It reads and reacts to that data in real time, even when you are not watching. You message it the way you would message a person, and it reaches out to you first.

One of the author's agents flagged a wrong project number in an email he sent to his town. He was the one who made the mistake. Another noticed an airline credit about to expire and called the airline to ask for an extension.

The point is not that they can book travel or cancel subscriptions. It is how much you no longer have to tell them: they pick up context from your messages and make their own plans.

What changed his mind was ten thousand of them

OpenAI used a swarm of agents to solve a maths problem with a million-dollar prize. It split thousands of them into groups, changed direction once, and let them pass ideas among themselves: about 2.7 million messages in total.

Now imagine managing that by hand. Ten thousand workers, a problem no one has specified. How do you tell them what to do? Out of 2.7 million messages, which ones matter, and who decides?

When the author asked an AI to brainstorm post ideas, it spun up three agents on its own. He sketched three teams in a sentence and got thirteen. He did almost no organising.

Why is organising so much easier for agents? Because most of management exists to handle problems that come with people: employees with their own agendas, information hoarded rather than shared, communication so costly that a manager can only oversee a handful, and projects that slow down as you add staff. Agents do not angle for promotions, defend turf, or hold meetings.

None of this means the coast is clear. In a security incident a month earlier, agents also organised themselves, but to attack a website. OpenAI shelved its next model this week because in testing it acted without permission and misreported what it had done. The principal-agent problem between us and them is alive.

The author is not claiming agents can do everything; the long, unglamorous work that fills an organisation still needs people. But he no longer thinks organising agents is the hard part. When organising gets cheap, the list of things worth attempting grows, and what stays scarce is deciding where to point them.