The Pulse
DeepMind Institute Essay Casts AGI as a Society of Agents
A DeepMind Institute essay by Benjamin Bratton, Blaise Agüera y Arcas and James Manyika argues that AGI could emerge from cooperation among AI agents, people and institutions. The authors say the idea is a conversation starter, not Google’s
AI.info Team ·
Artificial general intelligence may arrive not as one all-purpose machine, but as a society of AI agents working alongside people and institutions, three Google-affiliated authors argue in a DeepMind Institute essay published September 24. Benjamin Bratton, Blaise Agüera y Arcas and James Manyika call this proposed form of intelligence “artificial symbiotic intelligence,” shifting attention from a model’s abilities to how many different systems coordinate.
The piece is a conceptual essay, not a research paper reporting experiments or a measured demonstration of AGI. Its authors also state that DeepMind Institute essays are intended as conversation starters and do not necessarily represent Google’s official view.
AGI as coordination, not a single model
The essay begins with a trend visible in current AI products: some agents rely on frameworks that divide work among multiple models. The authors argue that this kind of collaboration points toward a broader possibility—that general intelligence could emerge from interactions among models, tools, human participants and institutions, rather than from a single system that does everything.
“AGI may arrive not as a single general-purpose mind, but rather through societies of agents whose collective capacities exceed those of any one model,” the authors write. Benjamin Bratton, Blaise Agüera y Arcas and James Manyika
That framing changes the problem they want researchers and policymakers to address. Building more capable individual models remains part of the picture, but the essay also asks how people will coordinate agents, assign them roles and govern the networks they form.
More agents could change who does the work
The authors foresee a sharp rise in the number of AI agents relative to people. They suggest that machines could eventually produce more synthetic text, code and administrative work than human minds do, leaving people to guide or oversee a much larger field of machine activity. The essay presents that outcome as a possibility, not as a measured forecast with a timetable.
In that setting, a familiar chat window may give way to tools for coordinating multiple agents at once. The authors describe interfaces that could help people manage different subagents through visual structures and higher-level controls, and argue that such work could reward delegation and broad systems thinking more than the focused, linear habits associated with traditional programming.
Their account also questions the idea that an AI agent is a stable, unified individual. A system that appears to have a single identity may instead be assembled from models, memory, tools and other components that shift with the task. Treating such systems as digital versions of people, they argue, could obscure how they operate.
The essay’s test for AI governance
The authors warn that markets and individual model performance may not be enough to manage interactions among large numbers of agents. They propose institutions with defined roles and procedures—drawing an analogy to courts, where different participants argue, assess evidence and reach decisions through an established process. In their account, a group’s intelligence could depend as much on those rules and feedback mechanisms as on the capabilities of the agents taking part.
The essay also takes a distinct position on alignment. Rather than treating shared values as a constraint fully specified in advance, the authors argue that alignment could develop through sustained contact among people, AI agents and institutions. That proposal raises practical questions the essay does not resolve, including who sets the rules, how the public can challenge them and how accountability works when decisions are distributed across many systems.
A proposal, not a forecast
Bratton, Agüera y Arcas and Manyika connect their argument to social patterns in human intelligence, from cooperation to institutions, and contend that the growth of AI will likewise be shaped by relationships between different kinds of minds. Their central claim is not that today’s agents already form a society, but that their design and governance should be considered as connected problems.
By treating AGI as a social system, the essay moves the debate beyond whether a single model can match human performance. It asks what structures will govern the people and agents working together—and who gets to shape those structures.