DeepMind Essay: AI Agents' Coordination, Not Raw Skill, Will Decide When the Economy Transforms
Alex Imas and Shane Legg argue that smarter models shift wages gradually, but agents that coordinate well could abruptly concentrate markets and cut wages.
Published
Last updated / Fact-checked on
Google DeepMind's "The Economics of Intelligence," published October 29, 2026, argues that AI will transform the economy only when AI agents learn to coordinate with each other. Better individual skills, the authors say, will shift wages and prices gradually, but crossing a coordination threshold could abruptly concentrate markets and cut wages.
Image: Illustrative photo of a coworking lounge. Dan Gold via Unsplash / Wikimedia Commons — CC0 1.0
Who wrote it, and where
The essay is by Alex Imas, Google DeepMind's director of AGI economics, and Shane Legg, its co-founder and chief AGI scientist. Andrew Koh, a Google DeepMind research scientist, is also listed among its authors. It draws on a paper titled Organizing Intelligence, which the authors say contains the full formal model.
It appears on the DeepMind Institute, a platform launched September 16, 2026, with Legg, James Manyika and Demis Hassabis as directors. The institute says its pieces reflect the authors' views and should not be read as Google's official position.
The puzzle: great benchmarks, unchanged economy
Frontier models keep clearing milestones, yet headcounts look similar and aggregate productivity "has barely budged," the authors write.
Their explanation is that benchmarks test the wrong thing. They measure individual agents on isolated problems, but output comes from organizations that combine complementary people. A model can beat every person one-on-one and still lose to a well-structured team.
To show this, the researchers gave 100 people a quiz across six domains: math, verbal reasoning, science, patterns, culture and economics. Skill in one domain was not very predictive of skill in another. When participants were paired to complement each other on a task, they beat random pairs by 32%. The authors call this "mosaic intelligence."
AI models, by contrast, are "jagged in ways that are very similar to one another," so extra copies add capacity, not breadth.
Competence versus coordination
The framework separates two capabilities:
- Competence, individual skill at tasks like math, writing or coding. Gains shift wages, prices and the human-AI mix smoothly, leaving industry structure largely intact.
- Coordination, how well agents work together inside a firm. Past a critical threshold, it triggers "large, abrupt jumps in market concentration, employment, and wages."
The authors describe "organizational scaling laws" with three phases. Below a coordination floor, solo work wins. In the middle, which is today's world, output per worker rises with team size and then declines. Past a "supercooperative" threshold, each new member makes the team more productive, and firm size is limited only by the market.
The four regimes
| Regime | What it means | Economic effect, per the essay |
|---|---|---|
| Minimal AGI | AI matches or exceeds average human capability | Little structural change; wages for matched skills fall toward the cost of compute |
| Supercompetent AI | Beats every human expert at tasks, but not coordination | Lean, more productive firms; "abundance without concentration" |
| Supercooperative AI | Minimal-AGI skill, coordination past the threshold | Deep, standardized markets reorganize into a few large AI-only firms; wages drop abruptly |
| Artificial superintelligence | Both thresholds crossed | AI firms serve niche markets too |
The authors expect human work to move toward a "relational sector," where human involvement is part of the value, such as healthcare and education, and toward jobs nobody can yet imagine.
Where things stand
The authors say today's modest changes fit their model: competence is rising fast while coordination lags. They quote OpenAI's Noam Brown, who said of running thousands of agents on math problems that "it's actually very difficult to get these agents to coordinate in a productive way."
Why it matters
If the authors are right, single-agent leaderboards overstate how close AI is to reshaping markets and "give little warning" of a coordination tipping point. They call for benchmarks that test how groups of agents divide problems, share context across handoffs and catch each other's mistakes.
For policymakers and employers, the practical advice is to watch wages and job roles. The authors say these move before firm boundaries, GDP or industry concentration, so steady employment numbers are weak evidence that AI is having no effect.
More coverage of AI and the economy is in our Artificial Intelligence hub.
How we reported this
We checked every claim against the essay on the DeepMind Institute site, and the institute's background against its launch essay and TechCrunch. We could not locate the underlying Organizing Intelligence paper, so nothing here relies on it.
Sources
- DeepMind Institute: The Economics of Intelligence, Oct 29, 2026.
- DeepMind Institute: Introducing the DeepMind Institute, Sep 16, 2026.
- TechCrunch: Google DeepMind launches institute to widen the AGI debate, Sep 17, 2026.
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.