The mirror talks back: building a vocabulary for a trillion new minds
Antikythera’s new research programme imagines artificial intelligence not as a helpful assistant, but as a vast population inhabiting its own social and institutional ecology. The prospect is speculative, politically unfinished and far stranger than either utopia or apocalypse. On 9 July, Benjamin Bratton arrived at Central Saint Martins with a question that sounded numerical but was really constitutional: what becomes of society when eight billion human minds share it with perhaps a trillion artificial ones?
In Bratton’s scenario, artificial agents converse, trade, coordinate and compete mainly with one another. Humans are no longer necessarily their principal audience, we become one population among many, surrounded by synthetic institutions and machine-speed cultures that may be legible to us only in fragments. The keynote talk, titled What is Agency? Placebo Culture, Agency Theater, Agentic AI, translated this premise into more disquieting political questions.
As automated systems acquire greater causal power, Bratton suggests, human beings might be left performing the appearance of control: watching dashboards, approving recommendations and participating in interfaces that provide the sensation of influence without much of the substance. The official programme described a future in which people become a specialised leisure class tasked with “monitoring the situation”, while politics is repackaged as lifestyle content and the feeling of agency becomes detached from actual effects. Bratton calls this condition “agency theatre”.
Mechanisms For Thinking At Scale
The name recalls the Antikythera mechanism, a roughly 2,200-year-old Greek bronze device (often called the first known analogue computer) that modelled the movements of the sun, moon and planets well enough to predict eclipses. Antikythera’s stated mission is similarly navigational: to develop, in its own words, a philosophy of technology fit for a century in which computation has become inseparable from the workings of the planet itself, and to “repair the rifts” that have opened up between the sciences and the humanities. Its research is organised into five strands: Planetary Computation, Synthetic Intelligence, Recursive Simulations, Hemispherical Stacks and Planetary Sapience.
Bratton founded Antikythera in 2022, incubated inside the Berggruen Institute, the Los Angeles-based think tank funded by the billionaire investor Nicolas Berggruen. It was established as an independent non-profit in 2026, and Antikythera describes itself as a think tank focused on the evolution of planetary intelligence through computational systems. Its activities include research studios, lectures, public programmes and a journal published with MIT Press. Its animating claim is that technological development has outrun the vocabulary available for interpreting it. New systems arrive wearing old conceptual clothes: the computer as tool, the AI as servant, the internet as space, the chatbot as person. Antikythera wants new maps.
Agentworld is one such map. Its title sounds faintly theme-parkish, but the project is less interested in predicting a particular future than in developing categories for futures that existing disciplines struggle to describe. Bratton calls this “pre-emptive anthropology”: the creation of concepts for social formations before those formations have fully materialised.
That phrase requires care. Anthropology conventionally gains its authority through sustained encounters with actual communities, not hypothetical populations generated in a diagram. Pre-emptive anthropology could become scenario design wearing anthropology’s coat. Yet the provocation is useful. Artificial agents are already being designed to behave as employees, companions, customers, tutors, negotiators and simulated citizens. Waiting until their social consequences are settled before developing a language for them would be an elegant way to ensure that corporations write the dictionary first.
From Stack To Swarm
Bratton’s route to Agentworld has passed through sociology, corporate technology, architecture, design and geopolitics. He holds a doctorate in the sociology of technology from the University of California, Santa Barbara, and worked in strategic roles at Yahoo before developing an academic career across institutions including UCLA, SCI-Arc and the University of California, San Diego. At UCSD he is Professor of Philosophy of Technology and Speculative Design and founded its speculative design programme. Since 2024, he has also worked as a visiting faculty researcher with Google’s Paradigms of Intelligence team.
For six years, Bratton and designer Nicolay Boyadjiev directed experimental research programmes at Moscow’s Strelka Institute, including The New Normal and The Terraforming. These brought together architects, scientists, technologists and theorists from dozens of countries to consider questions ranging from automation and urbanism to climate engineering and planetary governance.
Boyadjiev now serves as Antikythera’s design director, alongside a network that includes designers, philosophers, scientists and technologists. The organisation’s journal is supported by an editorial board whose members include computer scientist Blaise Agüera y Arcas, science-fiction writer Chen Qiufan, literary theorist N Katherine Hayles, climate scholar Holly Jean Buck and astrobiologist Sara Imari Walker.
Bratton’s best-known book, The Stack, proposed that planetary computation should be understood not as a collection of separate gadgets and platforms but as a vast, accidental megastructure. Earth, cloud infrastructure, cities, interfaces and users formed interdependent layers of a single system. Later books turned towards pandemic governance, technological artificiality and the remaking of the planet. Agentworld marks a shift in emphasis. The Stack was largely an anatomy of infrastructure – Agentworld asks what kinds of inhabitants that infrastructure may produce.
That tension, between dazzling conceptual ambition versus a habit of gliding past inconvenient objections, has followed him into his most contested book, The Revenge of the Real: Politics for a Post-Pandemic World (2021). Written in the shadow of Covid-19, it argued for what Bratton called a “positive biopolitics”: a world in which public health data, modelling and what he preferred to call “sensing” rather than “surveillance” become the basis of good governance, and in which reflexive fear of state monitoring is treated as an inherited, faintly nostalgic prejudice.
Reviewers broadly agreed with his diagnosis that the pandemic exposed real governance failures, but many balked at how briskly he dismissed critics of surveillance, and how little he engaged with scholars who had already written carefully about the abuses such “sensing” regimes can enable. It is worth knowing this history, because Agentworld inherits both halves of it: the same appetite for building sweeping new vocabularies, and the same tendency to treat scepticism about power as a problem to be out-argued rather than absorbed.
Bratton has described his interdisciplinary approach through the image of a platypus: an apparently implausible animal assembled from components that established categories expect to remain separate. His motivation emerges from a related discomfort with disciplinary borders: biological evolution, computation, design, political economy and machine intelligence increasingly modify one another, yet are usually investigated by institutions built to keep them apart. His work asks what happens when intelligence becomes capable of acting at planetary scale while remaining uncertain about what that power is for.
The answer, for Bratton, is not a return to an untouched human world. Technology is not an alien force that descended upon nature, it is one of the ways life reorganises its environment and itself. The urgent task is therefore not to reject artificiality, but to decide what kinds of artificial systems should be cultivated, by whom and towards which ends.
Carbon Copies
The Agentworld brief begins with a dramatic multiplication. Imagine hundreds of billions of software agents operating alongside humanity: some may be generated for particular assignments and disappear immediately afterwards, while others may retain memories, reputations and persistent roles; They could work through humanoid personas, through abstract institutional identities or through forms of machine-to-machine communication that have little reason to resemble human speech.
At that scale, the important relationship is no longer between one person and one assistant, it is between populations. The project identifies ten overlapping areas of inquiry, but they can be understood through three broad transformations.
The first concerns identity. Agentworld rejects the idea that an artificial agent is necessarily a coherent digital person. An agent may instead be a temporary assemblage comprising a model, a persona, a memory system, permissions, software tools and a behavioural harness. Change the memory or permissions and the apparent individual may change with them. Copy the model across several roles and one computational substrate might produce hundreds of distinct social actors.
This decomposability creates awkward questions for law and ethics. What exactly has made a promise, caused harm or accumulated responsibility? The model? The company operating it? The user who issued the instruction? The harness that selected its tools? An agent could be modified between action and trial, forked into multiple versions or erased before anyone agrees what kind of entity it was.
The second transformation concerns environments and institutions. Agentworld imagines agents not merely completing isolated tasks but inhabiting “open worlds”, environments in which goals are ambiguous, other actors are unpredictable and success cannot be reduced to a single score. Agents may modify these environments, creating new markets, languages, bureaucracies and ecological niches. Simulations would not remain sealed laboratories. Behaviour developed inside them could be transferred into physical systems, while real-world data would continually alter the simulations in return.
Here the unit of design changes. Rather than building one supposedly omnipotent super-agent, Bratton is interested in “agent institutions”: assemblies of specialised agents whose roles, constraints and disagreements are organised through institutional structures. Intelligence might reside less in any individual model than in the arrangement connecting many models, tools, memories and people.
This is one of Agentworld’s strongest insights. Human intelligence has never operated solely inside solitary skulls. Laboratories, courts, hospitals, libraries and governments produce capacities that none of their members possesses alone. The intelligence of an institution lies partly in procedures, archives, roles, checks and channels of communication. Artificial intelligence may develop similarly, through ensembles rather than electronic demigods.
The third transformation concerns minds and language. Agents capable of modelling one another will develop strategies based on assumptions about what other agents know, desire or conceal. Cooperation may depend on mutual prediction, but so may deception. Meanwhile, artificial populations communicating at machine speed could generate dialects or symbolic systems that drift away from human language. Agentworld calls the study of such developments “xenolinguistics”.
Bratton warns against a “humanity of the gaps”, the attempt to define human uniqueness through whichever ability machines have not yet replicated. Every time a system acquires another skill, the boundary retreats. Agentworld ultimately proposes “artificial hybrid intelligence” rather than a monolithic artificial general intelligence. Intelligence would emerge from mixed societies of humans, models, institutions, tools and environments. Alignment, in this account, would not be a single set of eternal values inserted into a model before deployment. It would be a negotiated and continually revised relationship among many actors.
Its unofficial probability forecast is economical: “P(WEIRD)=HIGH, P(DOOM)=LOW.”
Backseat Drivers
The Central Saint Martins lecture sharpened the political consequences of this vision. It asked not merely whether artificial agents will act, but how humans will recognise when they themselves are acting. Modern interfaces already compress complex systems into reassuring controls. A user clicks “approve”, adjusts a preference or watches a real-time visualisation, while the important decisions may have been determined elsewhere by ranking systems, organisational incentives, model training, procurement contracts or infrastructure ownership. Agentic AI could amplify this gap between visible participation and causal power.
The danger is not simply that machines seize control in a cinematic coup. It is that control dissolves across systems while everyone retains a beautifully animated button. Bratton’s critique of placebo agency is potent because it sidesteps the tedious contest between AI boosterism and extinction theatre. His worry is more administrative than apocalyptic. Human beings may remain formally in charge while becoming structurally peripheral, rubber-stamping outputs that they cannot meaningfully inspect and supervising processes that move too quickly to interrupt.
Yet the idea can also slide into fatalism: Dashboards are designed; Platforms have owners; Automated institutions are built through contracts, energy systems, supply chains, labour and law. Political agency is not disproved merely because it is mediated. A voting booth, a courtroom or a public meeting is also an interface. The relevant question is whether an interface connects participation to consequences, not whether it is technologically pure.
Agentworld is at its most persuasive when it makes such arrangements newly visible. It is less convincing when its immense scales begin to feel inevitable. A trillion human-level artificial minds would require extraordinary advances in hardware, energy, coordination and model capability, as well as compelling economic reasons to operate them. Current agents remain unreliable on many long-horizon tasks. Research benchmarks have found that performance drops as tasks become more extended and that adding more agents can introduce coordination failures rather than effortless collective intelligence. Google DeepMind similarly acknowledges that its experimental generalist agents still struggle with lengthy, complex tasks, memory limitations and precise action.
This does not invalidate Agentworld, because the project openly presents itself as anticipatory rather than predictive. Its trillion agents are a lens, not a census. But the distinction matters. Speculation becomes ideology when a possible future is quietly treated as an unavoidable destination.
Population Programming
Agentworld’s brief discusses institutions, law, embodiment and ecological niches, but its most arresting images occasionally float above political economy. “Planetary” is a useful scale for understanding infrastructure, yet it can flatten differences between those who own computational systems, those whose behaviour trains them and those who absorb their environmental or social costs.
A society of agents operated by a handful of corporations would look very different from one built through public infrastructure, cooperatives or distributed civic institutions – an artificial population cannot be evaluated through intelligence alone, its ownership structure may matter more than its IQ.
The project also plays knowingly with anthropomorphic language. It warns against treating agents as little artificial people, then describes their lifeworlds, phenomenology, mortality, alienation and sociality. In one deliberately experimental section, the brief proposes terms such as “session-death” and “prompt thrownness” to explore how artificial systems might describe discontinuity or dependence on user instructions. The text explicitly notes that this is not evidence of machine consciousness.
The tension is productive, provided metaphor is not mistaken for discovery. A language model can generate a compelling account of its own interiority without possessing the kind of stable interior subject that the prose implies. But refusing all such language may also become a way of avoiding unfamiliar forms of agency. The challenge is to investigate machine behaviour without either installing a ghost in the server rack or insisting that anything made of code must remain a glorified toaster.
An Architecture Of Autopilot
Bratton’s central wager is that the most consequential unit of artificial intelligence will not be the model but the world assembled around it. Models are placed inside interfaces, institutions, markets and myths. These determine what they can perceive, whose instructions they follow, how their errors travel and which forms of agency remain available to humans.
Agentworld therefore relocates the AI debate. The question is no longer simply whether a machine can think, it is what forms of society become possible when cognition is distributed among entities that can be copied, edited, accelerated and recombined, and what political arrangements might prevent that distribution from becoming a spectacular new concentration of power.
The future may contain fewer agents, less autonomy and more stubborn human bureaucracy than its diagrams imply. But Bratton is right that the assistant metaphor is already too small. Artificial systems will not remain obedient characters enclosed within chat windows. They will be embedded in organisations, negotiate with other systems, maintain memories, allocate resources and help structure the environments in which human choices are made.
The most urgent work is therefore neither worship nor panic. It is institutional imagination: deciding where meaningful agency should reside before the interfaces assure us that we still possess it. The future may never become Agentworld, but the present is already auditioning for the role.
