Three Google DeepMind authors argue that the economy will look familiar until AI agents learn to cooperate, and then it will change fast. Alex Imas, Andrew Koh and Shane Legg make the case in an essay on the DeepMind institute’s site, published on 29 October 2026 and drawn from their paper “Organizing Intelligence”.

Their first complaint is about how progress gets measured. Most benchmarks score one model solving a problem alone, but markets are won by organizations. To show why that matters, the authors quizzed 100 people across six subjects, from math and science to culture and economics. A person’s score in one area said little about their score in another. When the authors paired people to cover each other’s weak spots on a task, the pairs beat randomly matched pairs by 32%.

AI models, the authors say, are uneven in similar ways to one another. Adding a human specialist fills a gap in a team, while running a second copy of the same model buys more volume and no new range. So a model that beats the average person can still lose to a team of people who each cover what the others lack. The question that decides markets, in their telling, is whether a group of agents can beat a group of humans.

Two traits separate agents from workers. Copies of a model can be rented at the price of compute, so human pay in any task the model can match gets pulled toward that price. And the things that cap the size of human teams (thin attention, context lost in handoffs, clashing career incentives) need not apply to agents that share the same weights and can pass along a full working state. The authors point to the telegraph and the railroad, which reshaped production in the 19th century by speeding up handoffs rather than by improving any worker.

From there they split capability into competence and coordination. Better competence shifts wages, prices, and the human-to-agent mix gradually, with firms staying small. Coordination works differently. Past what they call the supercooperative threshold, each new agent makes a firm more productive instead of less, and company size is limited only by the market.

That yields four regimes. At minimal AGI, the first stage, wages fall for anyone whose skills an agent can match, while firm structure holds. With supercompetent AI, firms get leaner and goods get cheaper, without concentration. Supercooperative AI is the abrupt case: a handful of very large AI-only firms in big standardized markets, with pay dropping sharply, and displaced workers pushing down wages in niche and human-contact sectors. Artificial superintelligence removes the niche shelter too, leaving the authors to speculate that spending shifts toward work where a human’s presence is the product, such as healthcare and teaching.

Today’s evidence fits the first stages, they write. OpenAI’s Noam Brown, describing runs of thousands of agents on open math problems, said “it’s actually very difficult to get these agents to coordinate in a productive way.” The authors read the current labor market the same way: smaller startups, flatter org charts, weaker entry-level hiring, and steady demand for veterans who can steer large projects.

This is a theoretical model from a lab that builds the systems it describes, and the only data in the essay is the 100-person quiz. Nothing in it says where the supercooperative threshold sits or how close current agents are to it.

The authors do offer a way to find out. They call for benchmarks that score groups of agents on splitting unfamiliar problems, sharing context, and catching each other’s errors, and for statistics that track payroll alongside compute spending. Until those exist, a company planning headcount from single-model leaderboards is reading the wrong gauge, and the authors say wages and job roles will move well before headline GDP does.

Reported by the Google DeepMind institute (essay by Alex Imas, Andrew Koh and Shane Legg) on 29 October 2026.