The Economy That Nobody Is Watching: DeepMind and the Coming Agent Economy

The Market With No Floor

Imagine an exchange whose traders are all machines. (This is an imagined scene, not a description of any real market.) The floor is quiet. There are no shouts, no hand signals, no jackets. Thousands of small agents, each carrying a budget and a goal, bid for a few seconds of compute or a place in a queue. Prices move faster than a human can blink. Most trades settle before anyone reads the notification that they happened.

The old trading floor, the one in the photographs with its shouting crowds, was a place where people could be seen making money and losing it. The market I am imagining has no such window. It runs on screens, and the few who look at it are usually the people who built it.

This essay asks two questions about that quiet room. The first is whether a researcher at Google DeepMind has already described it. The second, which I care about more, is who counts as a citizen of such an economy: whose agent gets a budget, whose labour is priced, and whose door is left open.

A Warning Filed in September

On 12 September 2025, a seven-author paper called Virtual Agent Economies was submitted to arXiv. The lead author is Nenad Tomasev. The other six include Julian Jacobs, Iason Gabriel and Simon Osindero. The paper is on arXiv, and the summary below draws on its abstract.

Its central idea is the sandbox economy: a new layer of the economy where AI agents transact and coordinate with one another. The paper sorts such economies along two axes. The first is origin: did the economy emerge on its own, or was it designed on purpose? The second is permeability: how freely does the agent economy touch the human one? An impermeable sandbox is a room with a good wall. A permeable one is a room with a door onto the street.

The paper's diagnosis is blunt. On its current trajectory, it argues, the agent economy is likely to be spontaneous and highly permeable. That combination brings coordination benefits, and it also brings systemic and inequality risks. The proposed remedies are auction mechanisms for fair allocation, "mission economies" aimed at collective goals, and "steerable agent markets" built with trust, safety and accountability infrastructure.

A secondary summary from IndexBox adds a detail the abstract does not state: each user's agent would begin with an equal starting allowance of virtual currency, and the auctions could be inspired by distributive justice. That is a striking detail. It comes from the summary, not from the paper's own text, so treat it as the summary's account.

If the equal allowance is real, it is, in my judgment, a citizenship proposal written in the language of accounting. It gives each person a seat at the market before anyone asks whether that person can read the terms.

Judgment, offered as my own: the permeable sandbox is the default because walls are expensive and doors are cheap. Nobody has to decide to open the door. It opens every time a contract is signed across the boundary.

What DeepMind Has Actually Published

There is a premise to clear first. The phrase "DeepMind institute AGI economy" does not match a formal Google DeepMind institute with an AGI-economics mandate. What exists is a DeepMind Institute essay series at institute.deepmind.com. Its own disclaimer says the platform was started by researchers from Google and Google DeepMind, that the pieces are "conversation starters", and that they "should not be read as Google's official view." That is the most important sentence on the page, and it is easy to skip.

The work inside that series has a clear author. Alex Imas is bylined as Director of AGI Economics, Google DeepMind. In January 2026, Business Today reported that DeepMind was hiring a chief AGI economist, quoting Shane Legg's X post: "I'm currently looking to hire a Senior Economist, reporting directly to me." A famous AI scientist personally recruiting an economist is a small, telling detail. I read it as a sign that the question of what happens to an economy after AGI was, until recently, a question his field had not staffed.

Three documents matter most.

The first is "Economic Policy for AGI", dated 16 September 2026 and written by Julian Jacobs and Imas. It sets out a phased strategy with three scenarios. In mild disruption, unemployment insurance, the earned income tax credit and employer-led retraining scale up automatically. In moderate displacement, the credit becomes a negative income tax if prolonged unemployment and falling median wages appear. In structural decoupling, where labour's share falls while capital returns rise, a Universal Basic Capital backstop is prepared and deployed only under that condition. The essay evaluates eleven policies. The SSRN abstract linked from DeepMind's research page counts twenty-four interventions. This essay uses the essay's own count.

The surprise in that design is the trigger. The last stage is triggered by a condition rather than a date: labour's share falling while capital returns rise. The middle stage could be triggered if data shows prolonged unemployment spells and falling median wages.

The second is "The Economics of Intelligence", dated 29 October 2026, by Imas, Andrew Koh and Shane Legg. The third is "Organizing Intelligence", by the same three authors and dated the same day. Its footnoted disclaimer, in the paper's PDF, is careful in a way worth quoting. Its scenarios "are not empirical forecasts but outcomes based on the model's assumptions", and the paper "does not represent corporate forecasts, product roadmaps, or empirical statements regarding Google's current or future operations." This essay follows the paper's framing and does not discuss every result in its model.

Finally, the sandbox returns in 2026, though not as an economy. In June 2026, MIT Technology Review reported that DeepMind was funding research into what could go wrong when millions of agents interact. A $10 million fund is backed by Google DeepMind, Schmidt Sciences, ARIA, the Cooperative AI Foundation and Google.org. The article also says such research should be run in sandboxes, since LLM-based agents may not always act rationally. The same word appears, now as a laboratory rather than a market.

The phrase "sandbox economy" comes from the 2025 paper. Anyone who says DeepMind has launched a sandbox-economy programme is going beyond what that paper shows.

Four Doors Into the Same Room

The Imas, Koh and Legg essay lays out four regimes. Think of them as four doors into the same room.

In the first, minimal AGI, AI matches average human capability across tasks. Industries stay fragmented, the changes show up mainly in wages, and headcounts may understate what AI is doing. The second, supercompetent AI, beats every human expert at individual tasks but not at coordination. Firms become more productive without becoming larger. The third, which the essay calls supercooperative AI, is the one to watch. Standalone competence stays near minimal AGI, but agent coordination crosses a threshold where adding agents raises productivity. Deep markets can then reorganise abruptly into large AI-only firms, and wages in those markets drop abruptly. The fourth, ASI, exceeds both thresholds. Production concentrates in AI-driven firms, and human work moves toward relational sectors such as healthcare and education.

The "Organizing Intelligence" abstract makes a sharper claim. AI expertise can be reproduced, while human expertise is in fixed supply. Competence shifts the level of productivity. Coordination governs how productivity scales with the size of a firm. In its words, "as AI coordination improves, however, this can break the Coasean ceiling on firm size and the market structure undergoes a phase transition toward large, concentrated AI-heavy firms." In standard economics, the Coasean ceiling is the point at which a firm stops growing because organising more activity inside it costs more than buying that activity on the market. If coordination gets cheap enough, that ceiling moves, and the boundary between firm and market moves with it.

The wage point surprised me most. Wage effects, the paper argues, split into a technology effect and a reorganisation effect. A tool that makes each worker more productive can still lower wages if the firm is rebuilt around the tool. The machine is not the only thing that changes. The org chart does too.

"Reorganization can reverse the technology effect on wages."

These are model outputs, not forecasts, as the paper says.

Clocks and Ledgers

Some of the people building these systems will put dates on their guesses. What follows is what they have said, with the source for each.

The most recent public statements on timing come from Demis Hassabis. TechTimes reported on 22 October 2026 on his Wall Street Journal Magazine profile, quoting him as saying "It's soon" and "I'm updating nearer and nearer." TechTimes says the estimate is his shortest so far. The same report describes him calling AI-agent incidents early warning shots, and saying that safety infrastructure is not yet ready.

No official Google DeepMind document cited here gives a year for AGI. What exists instead is interview testimony from named individuals, and each carries its speaker's own odds. That is worth knowing before you put an interview sentence on a slide.

Now the labour market, where the evidence is measured rather than predicted. On 12 August 2026, the Stanford Digital Economy Lab published an update using ADP payroll data. For workers aged 22 to 25 in highly AI-exposed occupations, employment "now stands about 19% below where it would be" had it kept pace with similar workers in less-exposed occupations. The gap was 15 per cent at the July 2025 data vintage. The headline is the sentence to read carefully: "No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19%." Experienced workers show no comparable gap.

In the Stanford figures the gap sits on the first rung of the ladder: workers aged 22 to 25, while experienced workers show no comparable gap. In my judgment, that is how an economy can look stable from the middle while the entry points narrow.

The Last Room: Kenya at the End of the Pipe

Consider a Nairobi annotator, a single person. Imagine her at the end of a contract: a laptop open, a task list she can see but cannot change. (Imagined scene.) Her work is the labelling that makes machine learning systems usable, and it is the part of the pipeline that machine economics likes least to show.

The facts around her are real. On 16 April 2026, Samasource Impact Sourcing said it would lay off 1,108 workers in Kenya after Meta terminated a content moderation and data annotation contract. The report calls Nairobi a key hub for data annotation and content moderation work. One client's decision moved a whole labour pool. Judgment: this is the sandbox's permeability in human form, if the paper's framing holds. The wall between the agent economy and the human one is not built in a seminar. It is built, and unbuilt, in contracts.

Kenya's Special Envoy on Technology, Phillip Thigo, said on a Capital FM programme that about 170 million jobs could be created and 92 million lost by 2030, a net gain of about 78 million. The East African Herald reported this on 17 August 2026, and reports that the figures are said to be consistent with the World Economic Forum's Future of Jobs Report 2025. The sentence that deserves the room is the one he added: "The reality is jobs will be lost." And then: "Unfortunately, you cannot sugarcoat that."

Judgment, mine: I put more weight on the loss than on the net. A net figure depends on which jobs are counted, which tasks get automated first, and which new ones appear fast enough to employ the people displaced. The loss is what a person in Nairobi is likely to feel first.

Judgment, uncalibrated: the question for Kenya is not whether the agent economy will price its labour. It will, and it already does. The question is whether the people who train the machines will be counted as citizens of its market, or merely as the inputs it draws on.

Who, then, is a citizen of such an economy? The abstract's answer is auction mechanisms for fair allocation. A secondary summary adds an equal starting allowance for every agent, a detail the abstract itself does not state. The DeepMind essays answer more quietly. Their policy ladder begins with unemployment insurance, retraining and the earned income tax credit, and it reaches for a universal capital backstop only if labour's share falls while capital returns rise. That is a government's answer, and it presumes a government able to act. The annotator in my imagined scene may have an agent with an equal allowance, if the summary is right. She also has a contract that ends on a date someone else chose.

So citizenship in the coming economy is not a matter of who has an account. It is a matter of who holds the budget, who holds the latch on the door, and who is told that the room is "conversation starters" rather than policy. The DeepMind essay itself says it is not Google's official view. The word "official" is a door too.

The sandbox has a door. Somebody will decide whether it opens from the inside, and that decision is being made in contracts, such as the one Meta ended with Samasource, and it is also at stake in the auction mechanisms the paper proposes. Imagine the annotator at the end of her contract, looking at the one screen in the room that still shows her name. Then imagine that the price beside it is the only thing anyone asks about.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing (parallel Claude Haiku 5.5 agents) and checked against the linked sources on 30 October 2026. Convictions, proposals and forecasts are identified in the text; scenes framed as imagined are illustrative.

Cover photograph / illustration: Kevin Hutchinson (via Flickr, https://www.flickr.com/photos/hutchike/3880462147/). CC BY 2.0. Original source. Illustrative image: the trading floor of the New York Stock Exchange. Photograph: Kevin Hutchinson. CC BY 2.0. This image does not document the events described.