Yesterday afternoon I sat in a video call with a couple of dozen entrepreneurs. It is a small, invitation-only circle; everyone in it has built something, and once a month one of us shows the others what they are doing that nobody else is doing yet. This month the founder of an education company with close to two hundred employees shared his screen. The second window showed his MacBook. He said it was overheating.

He had rebuilt his company in two weeks.

Not restructured. Rebuilt. He described a swarm of forty-seven AI agents that talk to each other, one lead agent per product, a “queen” that orchestrates the rest, and a locally hosted dashboard he calls his headquarters: every agent, what it is working on, which agent is asking which for what, and a board of open decisions waiting for him. Eighty-nine buttons to click that evening. He said that was about a month of work.

The numbers he gave were his own and I could not audit them. His team ships one App Store release every two months; his agents produced twenty-one versions of an app in forty-eight hours. Two engineers had spent a year on a food-logging feature he was ashamed to launch; an agent built a better one in an hour, and he is replacing the year of code with the day of code. Ten thousand dollars in tokens in thirty days, four products, one of them now in soft launch. What I could verify is that the apps ran, in front of us, and that he does not read code. He said so plainly: he judges what he sees and directs. His rule for the whole operation was one sentence. People decide. Agents do the typing.

That rule is the part worth keeping, and I will come back to it.

What he actually said about the market

The other guest that evening teaches online courses and has reached well over half a million students on his own. He had flown in the night before. He asked the obvious question, the one I was asking silently: why forty-seven agents and not one? The answer was the most useful thing said all evening. Context. An agent asked to hold a learning platform and another app in its head at once starts making mistakes and has to compress its own memory. So you specialise them, exactly as you specialise people, and you let them talk. The organisational chart is back, and it is made of context windows.

Then the founder said something I think is right and most boards have not priced in. Ship in the fourth quarter, he told the room, because by the first and second quarter everyone else will be catching up. The barrier to building software is falling toward zero. A thirteen-year-old, he said, could now build the next version of his company.

I hold this view too, and I want to state it as a view, not as a finding. The moat in software was never the code; it was the ability to produce the code. When a founder can speak a working product into existence in an afternoon, what is left of that moat is audience, proprietary data, distribution and the judgment about what is worth building. Much of the SaaS landscape as we know it will become irrelevant for the same reason the typing pool did: the scarce skill stopped being scarce. If there is a product on your 2027 roadmap, the question is no longer whether you can build it. It is whether it has a reason to exist once anyone can.

The sentence that was not supposed to leave the room

In a different room a few weeks earlier, same kind of circle, same kind of founder, I heard the line that this whole movement is circling and will not say in public. Do not hire. Not a single full-time person right now, or you will regret it. And then, almost immediately: do not repeat that outside this room, it is a dangerous thing to say.

I am honouring the request by not saying who. I am not honouring it by pretending it was not said, because the sentence is the real content of the moment. The same founder, in the same breath, said he was discussing a four-day week with his executive team for next year. Two hundred people, and a founder who had just watched one of them, himself, outbuild his engineering teams fifty to one. He was not gloating. He was visibly unsure what to do with the people.

The evidence, for now, is more cautious than the mood in those rooms. McKinsey’s 2026 State of AI survey found only 14% of respondents reporting that AI had contributed to workforce reductions in the past year, less than half the share who had expected cuts a year earlier. That is a fact about last year. The rooms are talking about next year.

The flaw

Here is what nobody asked, and it has been bothering me since the call ended.

Wealth has only ever come from one place: something another person wants. A skill, a harvest, an hour of attention, a judgment. That is the entire basis of the exchange economy. Now run the founder’s logic forward, not for one company but for the eight billion. If most people no longer have anything to offer that anyone else wants, because the agents do the typing and the thinking and the designing, then where does the money come from with which they buy the app, the coaching subscription, the tokens?

The standard answer is universal basic income, and I think it is worth taking seriously rather than dismissing. But look at what it is. It is a transfer mechanism, not a source of wealth. Someone funds it. In this future the someone is the owner of the agents, and the owner of the agents is also the vendor the recipient will spend the money with. The customer is being financed by the supplier so that the supplier can have a customer. That is not an economy. That is a closed loop with a tax department in the middle.

The obvious objection is that every previous wave of automation created wants nobody had imagined, and the displaced found new things to offer. That is historically true and it is the strongest case against my worry. It is also the first wave in which the thing being automated is not a task but the general capacity to produce things people want. I do not know whether the objection holds this time. I notice that none of the founders in those rooms claimed it does; they moved to the four-day week instead, which is the quiet admission that the hours are no longer needed.

This is not reassuring.

Back to the rule

People decide. Agents do the typing. The founder meant it as a safety rule, and it is one. Read it again as an economic rule and it says something sharper: deciding is the one thing that still has to be done by a human, because it is the one thing another human will still pay for. The founder described his own job now as “to dream”, by which he meant deciding what gets built, for whom, and whether to say yes to the design board over his morning coffee. Everything else in his company has become a button.

For anyone running an organisation of two hundred, or two hundred thousand, that reframes the question entirely. The question is not how many people you will need. It is which decisions you are going to keep, who will be named on them, and whether you have started moving your people toward those decisions before the agents arrive at them first. A decision right nobody exercises does not stay with the person; it migrates to the machine. That migration is already visible on one founder’s overheating laptop.

Build the thing this quarter. Then decide what, in your company, is still worth a human’s name.