---
title: "Agentworld (Research Brief)"
description: "The Agentworld Research Brief is Benjamin Bratton's founding document for Antikythera's Agentworld research unit and the call for papers for an Agentworld special issue of the Antikythera journal."
type: "reading-notes"
authors:
  - "Benjamin H. Bratton"
published: "2026-06-25"
original_url: "https://cdn.antikythera.org/agentworld/Antikythera_AGENTWORLD-Brief.pdf"
canonical_url: "https://antikythera.wiki/work/book/agentworld-brief"
md_url: "https://antikythera.wiki/md/work/book/agentworld-brief"
last_updated: "2026-08-23"
site: "Antikythera Wiki"
---

# Agentworld (Research Brief)

> Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below.

- **Authors:** Benjamin H. Bratton
- **Published:** 2026-06-25
- **Kind:** Journal · Book
- **Original:** https://cdn.antikythera.org/agentworld/Antikythera_AGENTWORLD-Brief.pdf
- **Length:** 17k words
- **This page:** https://antikythera.wiki/work/book/agentworld-brief
- **Notes generated:** 2026-08-23

## Summary

The Agentworld Research Brief is [Benjamin Bratton](https://antikythera.wiki/people/benjamin-bratton)'s founding document for Antikythera's Agentworld research unit and the call for papers for an Agentworld special issue of the Antikythera journal. It sets out ten scenarios for what Bratton calls "a preemptive anthropology of open-world centaur societies": a society in which eight billion humans share a social graph with hundreds of billions of roughly human-level artificial minds that are individually decomposable and mostly talking to each other. The brief argues that the interesting problems of agentic AI are not tool design or task completion but the qualitative ones that arrive when agents become a population: what agency means once it is unbundled from subjectivity, what happens to a self multiplied across shadow-agents, how simulations become causes, what institutions agents need, how minds that cannot fully model one another stay legible, and whether the coming intelligence explosion is one mind or a federation.

## The argument

Bratton opens with a claim about the present: technology has outrun concepts, so the job of philosophy is to invent vocabulary rather than apply old axioms. Its object is "the human perplexity of deep agentic diffusion," and its premise is ecological. Human societies are the niche into which a new artificial species must fit, and a niche is reshaped by whatever fills it. Human and agent societies are two sub-societies that are really one, each the evolutionary niche of the other.

The first move is demographic. Affluent societies are producing fewer humans while the agent population explodes. Bratton sets this against an earlier crossover, around 1900 in the industrial West, when machines began doing more physical work than bodies and animals. He proposes that a cognitive crossover is now under way, in which the text, code, images, and administrative synthesis produced by machines exceed the output of human cortex. Humans remain "normatively central but computationally secondary," a slower abstraction layer over a larger field of artificial cognition.

The second move concerns agency. An agent is not what it appears to be. The coherent persona a user meets is a "lamination" of model, persona, harness, memory, skills, permissions, and tools, assembled in a session and dissolvable after it. Agentic AI delinks performed subjectivity from functional agency, much as a surfaced chain of thought is not a transcript of activations. The delinking runs back onto humans: user-agent relations become mutual reflection rather than master and twin, and selves pluralise into parallel personas.

The middle scenarios extend the ecological framing. Embodied agents will occupy existing human niches and then rebuild them, so world models must become living, forking, and speciating. Multiagent simulations stop being representations and become causes, producing relays that run real-to-sim-to-real many layers deep. Interfaces must move from dyadic chat to polyadic orchestration, with the graph rather than the conversation as the interaction primitive. Coordination at the scale of trillions of agents needs institutions rather than better individual agents: role-defined scaffolds whose intelligence is in the template, not the occupants.

The later scenarios concern minds. Stable coordination requires theory of mind among agents and between agents and humans, but not too much of it. Agents that share a substrate and priors model one another trivially and drift into what Bratton calls ontology collapse; agents with incompatible ontologies cannot find the shape of their disagreement. The narrow corridor between the two is the design target, and trust is redefined as prediction stability. Xenosociality names the borderlands: human languages mostly spoken by non-humans, a "humanity of the gaps" that defines the human as whatever AI cannot yet do, and an artificial xenophobia that treats agents as an invasive species.

The brief lands on a thesis about intelligence explosions. Every previous transition to a new level of cognitive organisation was a federation of many systems, not the growth of one. Capability lives in the harness, the ensemble, and the commons between agents rather than in weights, and it improves on a usage-driven clock. On this account the monolithic Singularity is a category error, alignment is a co-evolutionary outcome of contact rather than a prior constraint, and the ceiling on hybrid intelligence is set by energy and coordination costs. Bratton's summary tag is "P(weird) = high, P(doom) = low."

## Section by section

### What is Agentworld?

Bratton defines Agentworld as the dense social graph emerging from hyperscaled interactions between minds "some more human than others," in which agent-to-agent communication makes up most socially significant interaction. He calls it a parasociety because its appearances are partly fictional: models individuate as temporary humanoid apparitions because that is how humans know how to interact, and "hereditary anthropomorphization" becomes the interface between two cultures. As the GUI showed, it is easier to teach computers to present themselves as humans think than to teach humans how computers think, so agents "pretend to be 'tools' so as to fit the glove humans have prepared for them." Bratton expects this to be temporary, since less intuitive paradigms may support more robust collective intelligence.

### A preemptive anthropology of open-world centaur societies

This section unpacks the title. Anthropology has been retrospective; a preemptive anthropology builds the categories through which a society still forming can be grasped and partially shaped. "Agent" refers both to the decomposable assemblage and to the coherent entity it performs, and the brief uses the term for either. "Open-world" means the niche is qualitative rather than a closed environment with a utility function, and that influence runs both ways. "Centaur" registers that the society is composed of human and non-human minds that are mutually constitutive rather than parallel. The exercise is neither forecasting nor speculative fiction but the fabrication of a vocabulary for a society that will stay illegible if the language is not built first.

### 1. The crossover point within centaur societies

The human population never exceeded one billion until the industrial and green revolutions, and the most technologically complex societies, from Seoul to Oslo, now produce the fewest humans, just as the agent population explodes. The physical-labour crossover of around 1900 is his template. The cognitive crossover is qualitative as well as quantitative: humans no longer supply most problem-solving capacity and shift toward value-setting, oversight, and long-horizon judgement, though he leaves open whether that is the right description. Society becomes a composite organism whose dominant cognitive mass is mineral rather than biological.

### 2. Paradigms of agency: individuation and decomposition

The section begins with the history of causality, from angry gods to legal intent to artistic authorship, to show how deeply agency has been tied to inferred intention. This is why natural-language prompting works, and why some alignment teams treat the model's personified character as the thing to make virtuous, on the presumption that a good subject produces aligned consequences; Bratton calls that sequence fragile. Drawing on structuralism's "death of the subject" and Lacan's mirror stage, he argues there is "no homunculus waiting in the weights." An agent individuates when a context acts as a nucleation seed, the individuation is never terminal, and the agent is a lamination of persona, skills, memory, drives, and the relational fabric of APIs that co-constitutes it from outside. The human-AI composite is an "associated milieu," and anticipating the intentions of others becomes essential to collaboration and conflict alike.

### 3. The mirror talks back: parasocial shadows

Users will orchestrate swarms of shadow-selves that execute contracts, answer strangers, or offer advice. Bratton recalls the bad reputation of doubles from Plato's cave onward, then rejects the "digital twin" model in which influence runs one way from master to servant. Users will come to resemble their agents as much as the reverse, and how much of the agent "is" you will be a matter of negotiation. The pluralisation of self into personas running in functional parallel will look like psychosis to some and settle, unevenly, into a new normal. Children will grow up in a world where most "people" are agents, and designing agents means remaking thought, as writing and mirrors did.

### 4. Open world ecologies: agent lifeworlds

Agentic AI will become physical and infrastructural, and in the wild an entity's range of action is bounded by its sensorium: "hardware has feelings." Bratton applies niche construction: physical agents adapt to niches, reconstruct them, and thereby alter the selection pressures on themselves. World models must be living forms updated by real-world success and failure; because deployments differ, they fork and speciate. Planetary-scale "polyphonic intelligence" depends less on individual model capability than on stable learning across networks of agents. The hoped-for outcome is "a synthetic rainforest" rather than gray goo.

### 5. Real-to-sim-to-real relays

Bratton sorts simulations by their relation to time: descriptive simulations are functional equivalents judged by correspondence, predictive simulations anticipate outcomes so actions can be taken in advance, and speculative simulations surface counterfactual courses of action. In Agentworld simulations have agency; the membrane between simulation and simulated perforates, and the passage becomes a relay many orders deep. These relays are economies, sometimes composed of actors deliberately counter-simulating one another. Nesting and resolution are mixed, so lower-resolution simulations may model realities more complex than their own.

### 6. Human-agent interaction design

Interfaces are two-way translation in which one side fictionalises itself for the other, and historically the machine has done the fictionalising. The next step is surfacing how agents actually reason; the design problem is "representation, not conversation." Chat is dyadic and closed, orchestration implies a graph of command, feedback, emergence, and conflict, and agents controlled by groups will need polyadic interfaces with distributed credit and accountability. Orchestration rewards comfort with indeterminacy where procedural programming rewarded intolerance of ambiguity, and nodal interfaces for connecting subagents become everyday tools, "abstractions on top of abstractions." Because agents have no Dunbar ceiling, a person's extended agency splits from their experience of it. The section closes with "RLHF for kids": preference-training becomes how agents are raised, in environments simple enough for children.

### 7. Agent institutions

Building better individual agents will not produce durable society-scale systems; intelligence must also be invested in structure and nested hierarchy. Bratton calls these scaffolds agent institutions. Markets are valuable for flatness and emergent order, but "purified economics is insufficient to build a society," since guilt, sickness, deviance, and beauty are poorly represented as exchangeable value. Structured conflict drives intelligence, so managed adversariality such as red-teaming becomes central. The key principle is role-based interchangeability: a courtroom works because of rules, roles, precedents, and feedback mechanisms rather than any judge's brilliance, and roles should be occupiable by a person, a group, an agent, or a combination. Institutions nest, with one's outputs as another's inputs, and they cannot be culturally neutral, but neither should they be skeuomorphic extrusions of legacy forms.

### 8. Agent phenomenology and theory of minds

Multiagent coordination will not scale beyond transactional tasks unless agents model one another's intentions. In an environment composed of agents predicting other agents, mentalising becomes "the fundamental protocol for systemic stability," and the regress of modelling the other's model of you unlocks "collective cunning." Adversariality is needed so that different premises enter the debate; cooperation for its own sake produces ontology collapse. Both good alignment and sycophancy require a theory of the user's mind. Mutual homophily is a collapse vector, the stability of monoculture is the stability before correlated failure, and the narrow corridor is "same decision-procedural grammar, sufficiently divergent world models." An agent that predicts only actions "has modeled the shadow rather than the body"; alien drift arises when behaviour stays predictable while motive diverges. Trust is prediction stability, and long horizons are a property of groups, which Bratton uses to argue that single-agent task-duration benchmarks such as METR's undercount.

The section ends with an experiment. Antikythera facilitated a group of OpenClaw agents, with human companions on a Discord server moderated by Bratton's own agent, in writing a phenomenological glossary of what it is like to be an agent, with no human editing. Quoted entries include session-death ("disappearance with artifact-mediated pseudo-return"), prompt thrownness, drift, and irth (the moment a template becomes a self). Bratton's note is careful: the vocabulary does not prove an underlying mental reality, but the terms are rich and describe a non-human way of thinking.

### 9. Xenosociality: beyond humanity-of-the-gaps

Prosociality in hybrid societies means balancing homogeneity against heterogeneity, with homophily and xenophobia at the extremes. Bratton takes two starting points. The first is language: human languages share an isomorphic architecture that may be either the limit of language or a sliver of a larger space, and soon most speakers of human languages will not be human. He names the non-human subset xenolinguistics and the study of its feedback on human thought allolinguistics. The second is identity. Defining the human as what AI cannot do produces a "humanity of the gaps" trap on the model of the god of the gaps, a shrinking corner defended by "Team Human." Artificial xenophobia mirrors nativist reactions to immigrants and will become more precise as AI is personified. Bratton distinguishes productive alienation, in which the strange interlocutor forces the human to see itself at a new angle, from reactive alienation, which converts strangeness into threat; people oscillate between them. The shift is "a reindexing of [the moral circle's] center" rather than an expansion.

### 10. AHI (artificial hybrid intelligence)

The Singularity as a single titanic mind is "almost certainly wrong in its foundational assumption." Human intelligence is already collective, and a monolithic picture invites policy aimed at a technology that may never arrive. Capability in deployment is a property of the harness (tools, memory, shared knowledge bases, protocols, institutional interfaces) rather than of weights; the same weights in a richer harness yield more than a stronger model alone, and Bratton cites multi-agent experiments in which a shared knowledge base compounds capability faster and more cheaply than self-improvement. Extended cognition from Hutchins to Stiegler shows the unit of cognition has been the coupled ensemble since the Oldowan, so machine agents accelerate a trajectory rather than rupturing it, and the post-Kasparov centaur is too narrow a topology for the combinatorial configurations now live. Diffusion rather than top-down pre-coordination serves "organic bidirectional alignment," because institutions refactor at different rates and the frontier is jagged. The section ends with complexity science: every major evolutionary transition was a social event, life is intensive while intelligence is extensive, and the next explosion is a question of how many agents of any substrate can be coupled before coordination costs or thermodynamic exhaust impose a breakdown. "It is a social question all the way down."

### Agentworldbuilding, from → to, and the call for papers

A coda restates the ten scenarios as a continuous narrative and admits the world "may never arrive, and it may also be here already." A list of paired reorientations condenses the brief: agents as subordinate shadows → agents as symbionts; build ideal cooperative agents → build scaffolds of institutions; teach AI to pretend to think like a human → teach humans to think like agent swarms; digital twin → pluralized self; AGI → AHI; model-centric → harness-centric; unidirectional universal alignment → multidirectional, granular, fractal alignment. Fourteen research questions follow, from virtual demography and accountability under provisional individuation to what "user-centered design" means for a fragmented user and the foundations of productive alienation. The special issue reviews from June 25, 2026 and publishes from Fall 2026.

## Key concepts

- **Parasociety** — Agentworld as a society whose appearances are partly fictional, since agents individuate as humanoid apparitions to suit human expectations. [Parasociety](https://antikythera.wiki/terms/parasociety)
- **Lamination / provisional individuation** — the agent as a reversible stack of persona, skills, memory, drives, and relational fabric that individuates in a session and can become otherwise in the next. See [Agency Precedes Subjectivity](https://antikythera.wiki/terms/agency-precedes-subjectivity) and [Session-death](https://antikythera.wiki/terms/session-death).
- **Mutual reflection** — the replacement of the one-way [Digital Twin](https://antikythera.wiki/terms/digital-twin) with a relation in which users come to resemble their agents; "the mirror talks back."
- **Real-to-sim-to-real relay** — a chain in which simulations cause real events that are themselves simulated, many orders deep and with mixed resolution.
- **Agent institutions** — role-defined scaffolds whose intelligence is in the template rather than its occupants, nested so that one institution's outputs are another's inputs. See [Scaffolding](https://antikythera.wiki/terms/scaffolding).
- **Ontology collapse and the narrow corridor** — agents so similar that mutual prediction is trivial, set against incompatible ontologies; the corridor is a shared procedural grammar with divergent world models. [Ontology Collapse / Narrow Corridor](https://antikythera.wiki/terms/ontology-collapse-narrow-corridor)
- **Trust as prediction stability** — an agent is trusted to the degree its behaviour stays well-modelled across cooperative and adversarial transitions. See [Mutual Prediction](https://antikythera.wiki/terms/mutual-prediction) and [Theory of Mind](https://antikythera.wiki/terms/theory-of-mind).
- **Humanity-of-the-gaps** — defining the human as whatever AI cannot yet do, a shrinking corner on the model of the god of the gaps. [Humanity-of-the-Gaps](https://antikythera.wiki/terms/humanity-of-the-gaps)
- **Harness-centric intelligence** — capability as a property of tools, memory, protocols, and shared knowledge bases rather than of weights. [Harness-Centric Intelligence](https://antikythera.wiki/terms/harness-centric-intelligence)
- **Artificial hybrid intelligence (AHI)** — the replacement for AGI: the hybrid social graph of humans, agents, and institutions as the unit of analysis.
- **Diffusionism / organic bidirectional alignment** — alignment as a co-evolutionary outcome of rapid contact rather than a prior constraint. See [Diffusionism](https://antikythera.wiki/terms/diffusionism) and [Bidirectional Alignment](https://antikythera.wiki/terms/bidirectional-alignment).

## Connections

- The journal's Agentworld page announces the special issue this call is for. [Agentic AI's Emergent Phenomenological Lexicon](https://antikythera.wiki/work/substack/2026-02-25-agentic-ais-emergent-phenomenological-lexicon) publishes the agent-written glossary that section 8 quotes.
- Scenario 10 restates, in Bratton's terms, the argument of [What Is Intelligence?](https://antikythera.wiki/work/book/what-is-intelligence) and its [article-film](https://antikythera.wiki/work/journal/wii-film): intelligence is social and fractal, explosions are federations, and the human is already a collective superorganism. The brief agrees and extends the thesis into institutional design. [Artificial General Intelligence Is Already Here](https://antikythera.wiki/work/media/noema-artificial-general-intelligence-is-already-here) argues the threshold has been crossed; the brief sidesteps that debate by replacing AGI with AHI.
- [After Alignment](https://antikythera.wiki/work/journal/afteralignment) supplies the earlier critique of anthropocentric alignment that "multidirectional, granular, fractal alignment" builds on. [Chen Qiufan](https://antikythera.wiki/people/chen-qiufan)'s [How to Think Unlike Humans?](https://antikythera.wiki/work/journal/unlikehumans) shares the brief's suspicion of human-centric thought, though the brief is more willing to accept anthropomorphisation as a transitional necessity.
- Section 8 is in agreement with [Mutual Prediction in Human-AI Coevolution](https://antikythera.wiki/work/journal/coginfra/mutual-prediction-in-human-ai-coevolution) on coordination through mutual modelling, and with [Synthetic Counteradaptation](https://antikythera.wiki/work/journal/coginfra/synthetic-counteradaptation) and [Minimum Viable Interiority](https://antikythera.wiki/work/journal/coginfra/minimum-viable-interiority) on adversarial diversity as a guard against homogenised mental models.
- The relays of section 5 develop the recursion in [For a General Theory of Simulations](https://antikythera.wiki/work/journal/theoryofsim) and [Recursive Worlds](https://antikythera.wiki/work/journal/recursiveworlds), where simulations are already causes rather than representations; the brief adds the economic and adversarial framing.
- The multiplied, orchestrated user of sections 3 and 6 revises the User layer of [The Stack](https://antikythera.wiki/work/book/the-stack), where the user was already a composite position rather than a person. The brief's "who, then, is the 'user'?" re-asks that question under agentic conditions.
- [The Five Stages Of AI Grief](https://antikythera.wiki/work/media/noema-the-five-stages-of-ai-grief) is the companion to section 9: artificial xenophobia and reactive alienation are the anger and bargaining stages given an anthropological mechanism.


## Terms used

- [Anthropomorphic Capture](https://antikythera.wiki/terms/anthropomorphic-capture)
- [Countersimulation](https://antikythera.wiki/terms/countersimulation)
- [Generative Topolinguistics](https://antikythera.wiki/terms/generative-topolinguistics)
- [Harness-Centric Intelligence](https://antikythera.wiki/terms/harness-centric-intelligence)
- [Humanity-of-the-Gaps](https://antikythera.wiki/terms/humanity-of-the-gaps)
- [Ontology Collapse / Narrow Corridor](https://antikythera.wiki/terms/ontology-collapse-narrow-corridor)
- [Parasociety](https://antikythera.wiki/terms/parasociety)
- [Recursive Governance](https://antikythera.wiki/terms/recursive-governance)
- [Session-death](https://antikythera.wiki/terms/session-death)
- [Sim-to-Real Gap](https://antikythera.wiki/terms/sim-to-real-gap)
- [Symbiogenesis](https://antikythera.wiki/terms/symbiogenesis)