A familiar story about advanced AI imagines a single superintelligence improving itself beyond human control. Benjamin Bratton, Blaise Agüera y Arcas, and James Manyika are pointing to a different possibility: intelligence that grows through coordination among people, AI agents, institutions, tools, and shared rules.
Their Deepmind Institute essay calls this vision "Artificial Symbiotic Intelligence." It treats AGI less as a solitary machine breakthrough and more as a social system in which humans and machines shape each other over time.
From one superintelligence to many connected agents
The central claim is that artificial general intelligence, or AGI, may not appear as a single, self-contained entity. Instead, the authors argue it could emerge from systems where people and AI agents work together across many roles and tasks.
That view reflects how some of today's most capable AI systems already operate. Rather than relying on one model to do everything, they can divide work among several models and coordinate them as teams. The important unit is not just the model, but the network around it.
This changes the problem for AI research. If intelligence is distributed, then the challenge is not only to make a more powerful model. It is to coordinate and govern relationships among agents, people, and the infrastructure that connects them.
In this frame, intelligence becomes a social phenomenon. A system may reason, decide, and act through structured collaboration rather than through a single mind. That is why "Artificial Symbiotic Intelligence" directly challenges the classic singularity narrative.
Reasoning models show signs of internal plurality
The essay builds on work from the authors' wider research circle. One preprint, "Agentic AI and the next intelligence explosion," develops the social and institutional perspective behind the argument. Another, "Reasoning Models Generate Societies of Thought," looks at reasoning models such as DeepSeek-R1 and QwQ-32B.
In that second preprint, Junsol Kim, Shiyang Lai, Nino Scherrer, Blaise Agüera y Arcas, and James Evans examine reasoning traces. Their analysis suggests that these models can produce patterns resembling debate: they shift perspectives, raise objections, and reconcile competing approaches.
The striking point is that this behavior is described as emerging during training rather than being explicitly programmed. When reinforcement learning rewards reasoning accuracy, models may develop multi-perspective, conversational behavior on their own.
The Deepmind Institute essay extends that observation outward. If a single reasoning model can display something like an internal society of thought, then larger systems of people and agents may be designed as societies of cognition in a more literal sense.
AI agents are not fixed digital people
A major part of the argument depends on how the authors define an AI agent. They describe an agent as a temporary bundle of models, roles, memories, ethical orientations, tools, and skills. It may feel coherent to a user, but it is not anchored in the same way a person is.
A human self persists across interactions because the brain is a physically connected whole, even with internal division of labor. An AI agent, by contrast, is assembled again through each request. Its behavior depends on context, user intent, available tools, and the parts brought together for the task.
That makes the idea of a fixed AI personality misleading. The authors argue that treating agents as digital twins with stable human identities misses what they are and underestimates what coordinated agent swarms could become.
The same shift affects users. People may direct swarms of shadow selves to handle contracts or explore alternative identities. The essay calls these "parasocial mirrors" that talk back, suggesting that human subjectivity could become more plural as people work through many machine-mediated extensions of themselves.
Interfaces and skills may have to change
If AI work moves from one-on-one chat to coordination among many agents, today’s chatbot interface may be only a transitional stage. The authors imagine future interfaces that look more like visual network diagrams, where agents appear as nodes and users direct activity from an overview.
That would also change the skills people need. Traditional programming rewards sustained concentration, orderly steps, and low tolerance for ambiguity. Coordinating agent swarms requires comfort with systems that are harder to predict, along with the ability to experiment and delegate tasks to machines.
The essay also argues that people will need a theory of mind for machines. That does not mean assuming models have subjective experiences. It means learning how machine systems handle situations, break continuity, inherit context, and generate unusual states in their outputs.
The authors point to terms such as "Session-death" and "Prompt thrownness." "Session-death" refers to the end of a session as a break in continuity. "Prompt thrownness" describes being placed into a task with context already present, without having helped create it.
For the authors, these phrases are useful because they reveal differences between human and machine cognition. The goal is to understand agents without reflexively humanizing them.
Institutions become the real design problem
The essay’s most practical implication is that institutions may matter more than individual models. Better models and smoother interfaces are not enough if the central challenge is collaboration among many human and machine participants.
The authors argue that markets cannot manage every form of coordination. Basic social ideas such as guilt, illness, and virtue cannot simply be reduced to prices or transactions.
Instead, they call for institutions that define roles, rules, procedures, precedents, and feedback loops. The courtroom is their example: a judgment emerges through structured roles and ordered exchange between opposing sides.
Applied to AI, the point is clear. Agent systems need more than raw capability. They need settings that shape how agents act, how people supervise them, and how decisions are made.
Early signs already exist in orchestration harnesses, the control layers that coordinate several models. These systems can outperform individual models that are supposedly "smarter," which supports the broader claim that coordination can be as important as capability.
This also reframes AI alignment. Rather than imposing a fixed set of values on models from above, the authors see values as something formed through continuing contact among people, agents, and institutions. Since AI adoption moves at different speeds across fields, alignment becomes an ongoing negotiation rather than a one-time technical fix.
The result is a different picture of the future. The next major transition in intelligence may not be a lone machine crossing a threshold. It may be a growing web of human and synthetic cognition, organized through institutions that determine how that shared intelligence works.