Ten Years Out: The Last Decade of Execution Jobs

CLAIM: Not a forecast, an opinion: execution work has about ten years left, ownership becomes the dividing line, and the referee is a politics that has not yet noticed it is on the field.

This one is an opinion, and I want that flag planted before the first argument, because the opinion is more extreme than what I usually publish here. No forecasting model, no survey of experts, no careful retreat into "studies suggest". One engineer's honest view of the next ten years, dated June 2026, with a public grading scheduled for 2031 and 2036. Predictions age badly. That is exactly why they belong in writing, where everyone can watch.

It starts with a sentence from someone in a position to know.

A country of geniuses, on schedule

In October 2024, Dario Amodei, who runs Anthropic, published an essay called Machines of Loving Grace describing what is actually being built. Powerful AI, in his definition, is an AI "smarter than a Nobel Prize winner across most relevant fields", running not as one mind but as millions of copies that can work independently or all together on the same problem. His summary, in his words: "We could summarize this as a 'country of geniuses in a datacenter'."

Not a tool. A country. The wording matters because genius is, today, the scarcest input in the entire economy, and the one thing you cannot do with a human genius is copy-paste them. A country of geniuses that scales like software is the end of that scarcity, and economies reorganize around whatever just stopped being scarce. When computation stopped being scarce, every industry became a software industry.

I think this is close. Single-digit-years close. Amodei put his own date on it: "I think it could come as early as 2026, though there are also ways it could take much longer." From where I sit, in mid-2026, the ways it could take much longer are losing. My evidence is my own desk. In 2023 the best models autocompleted my code. In 2026 I describe a task in the morning, an AI agent, software that works toward a goal unattended, takes it from there, and I review finished output after lunch. The interesting part of my job has quietly shrunk to the describing and the reviewing. Extrapolating one more step like that is not science fiction. Refusing to extrapolate it is wishful thinking.

Intelligence grows hands

The standard comfort goes: fine, desk work is in trouble, but the physical world is safe. Plumbers, electricians, farmers. Software cannot fix a leak.

That comfort has an expiry date, and the reason is not a better robot arm. It is who designs the arm. Robotics was never really blocked on motors and grippers; it was blocked on two things. Machines were too stupid to handle the unstructured physical world, and hardware iteration was slow because every design cycle ran through a small number of expensive human engineers. A country of geniuses attacks both ends at once: it is the brain inside the robot, and it is the engineering department around the robot, simulating ten thousand design variants before anything gets machined, then designing the factory as well.

Once the design loop and the production loop both run on abundant intelligence, physical capability starts compounding the way software did. So, my opinion, stated plainly: within ten years, AIs design and produce robots for roughly any physical task that matters economically. Harvesters that cope with the terraced steepness of a Lavaux hillside. Home machines that do considerably more than vacuum. Construction sites where the foreman is the only one with a pulse.

Execution is what disappears

Here is the distinction that organizes everything I believe about the next decade. Every job mixes two ingredients. Execution: the goal is given, you perform the steps, and faster, cheaper, more reliable is simply better. Direction: you choose the goal, you own the outcome, you answer for it when it goes wrong.

The dividing line of the next decade is not human versus machine. It is execution versus direction.

Execution is precisely what a country of geniuses does better, around the clock, at near-zero marginal cost. So execution goes first, roughly in order of how digital it is. Executing desk work is already going: report writing, claims processing, first-line support, bookkeeping, most coding to spec, which is writing software to someone else's instructions. Soon, I think, essentially all of it. Physical execution follows as the robots arrive. Ten years out, my call is that every job whose core is "perform the defined task" is done better by something that is not a person.

What survives longer is direction: strategy, oversight, accountability. Not because machines will be unable to think strategically, but because someone has to choose what is worth doing and stand behind the result, and no legal system or customer on earth is ready to point at software when it goes wrong. Though longer is not forever. Even that reluctance can be priced and insured away, and one of the endings below does exactly that.

The uncomfortable arithmetic is how much of today's employment is execution. Open any national employment table: the biggest categories are retail, transport, logistics, food service, clerical work, machine operation. Defined tasks, almost all the way down. No insult to anyone's work; jobs have mostly been exactly that since jobs existed.

The years are guesses and the list is unfair to a hundred professions I left out. The shape is the argument: read it as water rising through the economy, lowest and most digital floors first. The hollow dot is a fade, not an ending. The lines that end all end for the same reason, and the lines that keep running all keep running for the same reason.

The ownership cliff

Now the part that actually worries me, because it is not about technology at all.

Selling labor is how nearly everyone gets their share of the economy. And the price of labor rests on one assumption so old it has turned invisible:

wage   = f(scarcity(human_labor))    # pay tracks how scarce human labor is
scarcity(human_labor) → 0            # what full automation does
wage   → 0                           # the channel closes
share  = ?                           # the line politics must rewrite

Wages are not a law of nature. They are a distribution mechanism, and the mechanism assumes human labor is scarce. Remove the scarcity and wages do not gracefully adapt; the channel simply closes.

Meanwhile the other channel keeps working perfectly. A fully autonomous service has revenue and almost no payroll. Its proceeds flow to whoever owns it and get reinvested into more autonomous capacity, and with no payroll there is no leak through which the money reaches anyone outside the loop. Ownership compounds; it does not spread on its own. Follow both lines for ten years and you reach the scenario I consider the default if nothing intervenes: a small class that owns highly autonomous services, and a very large class holding nothing the market still wants to buy. Not because anyone got lazy. And retraining does not fix it, because the job you retrain into is also execution.

A few endings, all of them extreme on purpose

Scenarios are flashlights, not forecasts; you point them at the dark corners deliberately. Here are three I can reach from today without breaking anything we currently know; none of them is scheduled.

The estate. By 2036 a few thousand people own the autonomous services everyone else lives on. Their companies need no employees: agents run the operations, robots run the sites, the lawyers are software and the software carries its own malpractice insurance. And the owners are, this is the uncomfortable part, nice. They fund hospitals, match donations, give thoughtful interviews about responsibility. The transfers they pay keep every fridge full. It is still the widest wealth gap in human history, administered with perfect manners: on one side compounding ownership of everything that works, on the other a livelihood that is no longer a right but a favor. There is an old word for an economy where the many live on the goodwill of a generous few, and the word is not "innovative". It is feudalism with better interfaces.

The dividend. Around 2030, some government does the boring thing on time: it taxes autonomous production at the source, the way Norway taxed North Sea oil, then does the thing Norway never did and wires every citizen a dividend. Work becomes optional within a decade. The crisis that follows is real but brand new: not income but meaning, status, what a Tuesday is for. The first crisis in history that arrives with the rent already paid.

The ballot. The owners have the economy, but the workless majority still has the vote, and in a democracy those are two different kinds of power. Somewhere in the early 2030s a government gets elected on a single promise: take the autonomous services into public hands. Maybe it ends like Switzerland would do it, a sovereign fund with every citizen a shareholder. Maybe it ends with capital fleeing to friendlier jurisdictions overnight and the lights flickering. The point is that the gap does not stay quiet just because the fridges are full. Sooner or later, ownership gets put on a ballot, and nobody has written the question yet.

An appeal to a politics ten years behind

Which ending we get is not a technology question; the technology is roughly identical in all three. It is a politics question, and politics keeps a different clock. GDPR, Europe's data-privacy law, was proposed in 2012 and enforced in 2018, six years for a problem that was already a decade old when drafting began. Call it six to ten years, in a healthy democracy, from "we should do something" to "something is enforceable". Now hold that interval next to the timeline above.

Parliaments in 2026 are debating retraining programs and four-day weeks. Those are answers to the 2010s. The question actually on the table is who owns autonomous production, and through which channel its proceeds reach people who no longer have labor to sell. The instruments that question needs are boring and slow to build: a tax base that does not assume payroll, transparency about who owns the autonomous services, dividend plumbing built and tested before the emergency rather than during it. Started in 2026, they are ready in the early 2030s, roughly when the waterline reaches the building sites. Started when the unemployment statistics force the issue, they arrive after the estate has built its walls.

The worst ending does not need anyone to vote for it. It only needs a politics that keeps answering the previous decade's question.

Signed, dated, flagged extreme

I called this an opinion at the start and I will flag it once more here, because I know exactly how it reads. The middle of an exponential is the worst seat in the house for judging one: everything behind the dot in the chart below looks flat, everything ahead looks vertical, and I am guessing about where the curve turns like everyone else.

So discount me accordingly, but hold me to it: this essay gets graded here, in public, in June 2031 and June 2036. If execution work is still the backbone of employment in 2036, the grading will be short and embarrassing, and I will write it anyway. If it is not, the grading will be the least of anyone's concerns. Either way, you read it here, dated and signed.