A second working population is arriving.

Every serious conversation about artificial intelligence is a conversation about capability. Parameters. Compute. Benchmark scores. Context windows. How well the thing reasons, and how much better it reasoned this quarter than last.

These are real measurements of a real thing. They predict what a system can do.

They do not predict what it will produce.

The gap has been widening even as capability improves, which is the first clue that capability is not the constraint. Something else is holding the door shut.

To find it, stop asking what an intelligence knows. Ask what it is permitted to do.

Picture a system that genuinely works. Not a demo. It drafts a commercial contract more cleanly than a second-year associate. It reconciles a messy set of accounts without complaint. It researches a market for eleven hours without losing the thread.

By any measure of capability, it is employable.

So let us try to employ it.

Every one of those six refusals is an institutional fact, not a technical one. Not a single one would be fixed by a better model. You could double the intelligence tomorrow and every answer would still be no.

In 1494, a Franciscan friar named Luca Pacioli published a textbook containing a section on bookkeeping. It described a method for recording every transaction twice, once as a debit and once as a credit, so that the books could be checked against themselves.

Nobody became more intelligent in 1494. But for the first time, a merchant in Venice could trust the accounts of a partner in Bruges he would never meet. Trust became portable, and trade followed it across a continent.

The pattern repeats. Consider what actually unlocked economic output:

Not one of these inventions made a single human being smarter. Each of them made millions of human beings more productive, by changing what a person was permitted to do with the intelligence they already had.

Which reframes the question entirely. The interesting problem is not how to make these systems more capable. It is how to make them citizens.

Strip away the abstraction and economic participation reduces to five permissions. A human acquires them so gradually, and so early, that they become invisible. An artificial agent has none of them.

Which is why the missing piece is not a better agent. It is the fabric an agent operates inside. And that fabric does not exist yet, for the same reason double-entry bookkeeping did not exist in 1493. Nobody had needed it before.

Suppose an agent receives all five. It has standing, an account, spending authority within limits, a reputation that travels, and access to capital. It does useful work and it is paid.

Now watch what it does with the money.

It buys its own compute, which extends how long it can work and how hard it can think. It hires other agents for capabilities it lacks, which means its earnings become another agent's revenue. And with access to capital, it can finance entirely new agents against expected output.

For the first time, labor supply can be manufactured out of retained earnings. Not hired. Not trained. Manufactured, by the labor that came before it.

The consequence is that output stops scaling with headcount and starts compounding. That is a different kind of curve, and it arrives whether or not anyone has prepared for it.

Take the second path seriously for a moment. Suppose ownership of productive agents really is spread widely, held by ordinary people who did not build any of it.

What have those people become?

Not customers. A customer buys output and leaves. Not employees, since the work is not theirs to do. Not quite investors either, because the thing they hold is not passive: it acts on their behalf, continuously, and its interests can be aligned with theirs by construction rather than by promise.

What they have become is members.

This sounds grander than it is. The requirements for a society are unglamorous and few. It needs a shared purpose, so that action has a direction. It needs persistent membership, so that people are still there tomorrow. And it needs the capacity to act collectively, so that the membership can do something the members could not do alone.

That is the whole list. Notice what is not on it. No territory. No government. No currency. No flag, no recognition, no seat anywhere. Those are implementation details that arrive later, and arrive precisely because they became necessary rather than because someone declared them.

Which means this thing can be real long before it looks like anything. And it explains why the pieces that matter are so mundane. Membership needs a record. Collective action needs a way to decide and a way to spend. Shared purpose needs to survive disagreement, which is the only real test of it.

Rails become jurisdictions the same way. Not by declaration, but by accumulation. When enough consequential activity has to pass through the same place, that place starts behaving like a jurisdiction whether or not anyone intended it. It gains the ability to set terms, and with it the obligation to be trustworthy, because everyone inside it is now exposed to its decisions.

The membership had three conditions, and none of them was a body.

Everything so far assumed a screen. Contracts drafted, books reconciled, capital allocated. Real work, but not the part of the economy you can stand in.

Most of the world is not information. It is objects that need moving and hands that have to be somewhere at a particular hour. Machines are getting good at that, and the bodies to carry them are getting cheap.

Which means the threshold was never intelligence.

It is an hourly rate. A machine does not have to be brilliant. It has to cost less per hour than the hour it replaces, and the cheap platforms are already being argued below thirty dollars an hour on the dull tasks.

Identity. An account. Bounded authority. A reputation. Financing against the work it will do. A robot that works a shift has exactly the same missing paperwork as an agent that drafts a contract, because institutions do not care what the worker is made of.

So nothing above needs amending. It just stops being about software workers and becomes about everyone who works. Both collars, the drafting and the lifting, taken by the same kind of actor for the same reason.

With one addition.

Software could be written by anyone with a laptop. A body has to be built. Actuators, factories, models, deployment data, the years of learning that come from machines already working. That is not a purchase. It is a capability, and it takes a long time to acquire.

Whoever holds it holds the title to every hour those machines ever work.

Which makes the last institution on the list the ability to build physical intelligence itself, held by the members rather than assembled quietly by whoever gets there first.

Because it will be assembled either way. Either a handful of firms own the means of physical production outright and everyone else is offered retraining, or the people who work beside the machines are on the title.

The second one looks smaller and more ordinary than people expect. A town in daylight. Plants, warm materials, useful activity, humans and machines in the same square, and nobody in it who has to ask permission to earn.

That is future tense, and it should stay future tense until the hardware has earned it. But it is the only end state worth aiming at, and there is no route to it that starts after the capability has already been concentrated.

Follow the argument to its end and it produces a specification. Not a product idea. A list of things that must exist before any of the above is possible.

It is a strange specification. It is not a model company, and it is not a wallet. It is closer to a civil service for an economy that does not exist yet, built in advance of the citizens who will need it.

It is also unglamorous. There is no benchmark to win. The work is identity plumbing, settlement rails, reputation records, capital structure, and the deeply boring discipline of publishing your own unflattering numbers.