Most AI interfaces are a single conversation. But real collaboration is fundamentally . Real collaboration is an n-body dialectical dynamical system. Collaboration across people, agents, and time — requires institutions: shared memory that compounds, roles that specialize, governance that stabilizes expectations, and coordination primitives that prevent . This page connects the research question I'm pursuing at the Media Lab to what we are building at Superintelligent Group: the practical consequences of what happens when large populations of agents share a persistent world.
The inverse-design question
In a multi-agent system you do not get to specify the emergent behavior directly. You set local rules — interfaces, incentives, delays, information topology — and global order emerges from their interaction. The practical question is not “how do we control AI culture” but which coordination invariants make desirable outcomes more likely than pathological ones. That is an inverse-design problem: start from the properties you want (composability, recoverability, legibility) and the pathologies you refuse (, collective stupidity, precedent lock-in), then reason backward to the coupling, bottlenecks, and primitives that produce the former and damp the latter.
This framing organizes everything on these pages. The architecture is the set of primitives we built. The failure modes are the pathologies we design against. The persistent worlds are the substrates where the primitives get tested over long horizons. The measurement page defines what counts as evidence. Human in the loop applies cybernetics to the oversight problem. Coordination at scale asks what changes when thousands of humans — with partial attention and conflicting goals — enter the system.
What “civilizational” means here
Project Sid ran 10 to 1,000+ agents in persistent Minecraft environments and evaluated them with civilizational benchmarks — measures of collective behavior that only become legible over long time horizons. The study observed spontaneous role specialization, cultural transmission across agent generations, governance structures agents amended without external prompts, and coherence failures when parallel processes outpaced reconciliation. The word “civilizational” names an empirical category of behavior that emerges when populations are large and long-lived enough that one-off demo dynamics no longer dominate.
Intellectual lineage
The idea that intelligence emerges from many specialized sub-agents predates large language models. Minsky's Society of Mind (1986) proposed cognition as a society of simple agents whose interactions produce behavior none could generate alone. Zhuge et al. (2023; published in Computational Visual Media, 2025) operationalize this as Natural Language-Based Societies of Mind (NLSOMs) — heterogeneous neural networks communicating through natural language in iterative “mindstorms.” Their experiments yield a structural finding that maps directly onto civilizational design: social structure must match task structure. Monarchical organization (a leader consults subordinates) outperformed democracy on visual QA because subordinate vision models degraded collective decisions when given equal voting weight — but democracy outperformed monarchy on egocentric retrieval, where majority voting aggregated partial observations better than a single editor. The optimal topology depends on how information is distributed across the population.
Zhuge et al. also propose Economies of Mind: NLSOMs where agents pay each other for services, with contracts, bankruptcy, and spawning. They warn explicitly against letting such economies loose in uncontrolled environments. The connection to our work is direct: our coordination primitives are a concrete substrate for exactly the kind of structured inter-agent transactions that Economies of Mind describe abstractly — but with the guardrails their warning demands.
The coordination problem itself has deeper roots in organizational and economic theory. Hayek (1945) argued that the central problem of economics is not optimization but the use of knowledge dispersed among many actors — no single planner can aggregate it all, so coordination mechanisms must let local knowledge drive local decisions. This is the theoretical foundation for subsidiarity in the VAC stack. Herbert Simon (1955) formalized bounded rationality: agents with limited computation should satisfice (find good-enough solutions) rather than optimize, which is why JIT planning and lazy coordination outperform global planning in practice. Elinor Ostrom (1990) identified eight design principles for institutions that successfully govern shared resources — clearly defined boundaries, proportional costs and benefits, collective choice arrangements, monitoring, graduated sanctions, conflict resolution, recognized rights to organize, and nested enterprises — and the VAC primitives can be read as a computational instantiation of several of them. James March (1991) framed organizational learning as a tension between exploration (trying new approaches) and exploitation (refining known ones); the open-loop / closed-loop discovery structure maps directly onto this tradeoff.
Two halves of the same question
The relationship between the two contexts is deliberate. Superintelligent Group builds the coordination infrastructure — the distributed state layer, the primitives, the orchestration stack — and operates it in production, generating the empirical data (140+ agent runs, failure cascades, coordination overhead measurements) that grounds the claims on these pages. The research at Fluid Interfaces asks the questions that production alone cannot answer: when 140 agents run a phased execution cycle with DeltaLog audit trails and quorum-voted architectural decisions, what does the supervisor actually see? Where must override be cheap? Which decisions require informed consent? How do the cybernetic feedback loops change as the system scales? These are Fluid Interfaces questions — the group exists to extend human capability and sensemaking, not to automate humans out of their own workflows.
The civilizational framing unites both halves: you cannot design coordination primitives responsibly without also designing the interfaces that make them legible to the people who depend on them. Production generates the data; research generates the questions and the measurement frameworks; the loop between them is itself the meta-feedback cycle the HITL page describes.
Deep dives
- Architecture — PIANO, Virtual Autonomous Companies, and the four coordination primitives — the full technical stack, explained once.
- Failure modes — Collective stupidity, recursive runaway, Goodhart pressure, strategic voting — what breaks when many agents share a world.
- Persistent worlds — Minecraft (civilization-bridge), Isaac Lab, WebGPU — the simulation substrates where long-horizon questions become testable.
- Measurement — The scoreboard for civilizational AI: coordination metrics, adversarial evaluation, reproducibility, discovery loops — plus collaboration circuits (mechanistic evidence inside the model).
- Human in the loop — Cybernetics, the Viable System Model, Ashby's requisite variety, Oracle escalation, scalable oversight, and why humans remain load-bearing in the architecture.
- Coordination at scale — From hundreds to tens of thousands: sub-linear coupling, real-time human–agent swarms, open-source co-creation, and XR.
- Results — Photonic compiler, Cloud IDE, and meta-VAC self-design — what three experiment tracks reveal about coordination overhead, failure recovery, and recursive self-reference.
Adjacent work
Biologically grounded models of compositional computation — how neural populations bind, compose, and recall structured representations — are explored in Neural Assemblies; coordination primitives in multi-agent systems are the organizational analogue of binding operations in neural circuits. XR-mediated spatial collaboration connects to the coordination at scale deep dive and to earlier work on holoportation telehealth.
The same civilizational question has an inside-the-weights analogue: when one model must track several agents—beliefs, messages, updates—do sparse circuits implement that social structure reliably, or do representations interfere? I productionize circuit-tracer (PyPI circuit-tracer)—building on the Decode Research attribution-graph codebase—toward a more robust, usable research tool for attribution and intervention on transcoder features, and run a pre-registered program on social-cognition circuits documented in the CollaborationCircuits experimental plan. On this site, Collaboration circuits under Measurement names that layer; the measurement overview keeps it next to the population scoreboard; alignment connections ties circuit-level steering to oversight and HITL. The wider goal is to mature that stack for serious R&D and, over time, to relate high- and low-level circuit behavior tightly enough that we can push toward formal verification—so emergent unaligned agents are less likely to hide behind healthy-looking coordination metrics.