We propose Holon — a distributed multi-agent architecture modeled on the structural and functional principles of biological neural networks. Rather than scaling intelligence by building larger models, Holon scales by multiplying complete, autonomous agents that self-organize through local neighbor communication, form persistent pathways, and produce emergent collective intelligence. The result is not a smarter model. It is a smarter architecture — one that gets more intelligent the more it runs, not just the more it is trained.

Not a Brain. A Society.

The dominant metaphor for AI at scale is the brain — a central intelligence that grows larger, deeper, and more capable. We reject this metaphor entirely.

A brain is one organism. It has one purpose, one substrate, one convergence point. It is designed to produce coherent, consistent, predictable output. Chaos is a failure state for a brain.

A society is something else. It has no center. Agents within it have their own goals, their own relationships, their own histories. They cooperate and compete, specialize and generalize, form alliances and dissolve them. No single agent controls the outcome. The output is emergent social behavior — and it is unpredictable by design.

Holon is not a brain. It is a synthetic society of AI agents — each a sovereign Holon, each operating with partial information and local relationships, producing collective intelligence that no single agent could generate alone.

1.1 Game Theory: The Price of Anarchy and the Price of Stability

Game theory gives us precise tools to reason about societies of self-interested agents.

The Price of Anarchy(Koutsoupias & Papadimitriou, 1999) measures how much collective performance degrades when agents act selfishly rather than in coordination. The Price of Stability measures how much the best stable equilibrium falls short of the true optimum.

In Holon, these are measurable system properties:

• ·Price of Anarchy — how much worse is the ensemble output of independently-acting Holons vs. perfectly coordinated inference? The regression layer minimizes this gap.
• ·Price of Stability — how close is the equilibrium the network self-organizes toward to the theoretically optimal network topology?

maximize(emergent collective intelligence) subject to minimize(catastrophic divergence)

Too much coordination kills creativity. Too little kills coherence. The art of Holon architecture is tuning that ratio — knowing when to let the society be chaotic, and when to let the regression layer impose order.

The Holon: Whole and Part Simultaneously

The term Holon was coined by Arthur Koestler in The Ghost in the Machine (1967), derived from the Greek ὅλος (holos, "whole") and the suffix -on (particle/part, as in proton, neutron).

A Holon is simultaneously a complete whole — self-contained, functional, capable of independent operation — and a part of a larger whole — contributing to a system greater than itself.

Each agent in the Holon network is:

• ·A complete AI instance with its own memory, tools, skills, and identity
• ·A specialist in a narrow domain (signals, content, analysis, outreach, reasoning)
• ·Connected to a small set of neighbors (typically 8–12)
• ·Capable of firing, receiving signal, passing output, and going dormant

Architecture: Two Minds, One Brain

Daniel Kahneman's Thinking, Fast and Slow (2011) describes two systems of human cognition. Holon implements this architecture directly, mapped to two communication protocols.

This is not an engineering convenience. It is a cognitive architecture. Holon thinks the way minds think.

3.1 MQTT — The Fast Brain

MQTT is a lightweight publish/subscribe protocol designed for real-time sensor networks. The analogy to Holons is precise: a Holon is a sensor — a specialist firing a signal based on its domain. Sub-10ms latency. Wildcard subscriptions. QoS levels for critical vs. fire-and-forget signals.

3.2 Email — The Slow Brain

Each Holon has a real email address at @intervals.so. The inbox is the agent's consciousness. Its sent mail is its memory. Its threads are its relationships.

This maps to the Actor Model in distributed systems — but implemented on real email infrastructure, making every agent's inner life human-inspectable by default.

3.3 Neural Pathways: Local Topology, Global Intelligence

No Holon knows about the full network. It knows only its neighbors. Signal propagates through the network topology. The topology is not fixed — pathways that produce correct outcomes are reinforced. Pathways that go cold are pruned. The network learns its own structure over time.

3.4 The Voting Brain

For immediate convergence, a query is broadcast to a relevant cluster simultaneously. Each Holon fires independently. A learned regression function maps the distribution of outputs → one coherent answer. Holons that consistently contribute to correct answers receive higher weights. The regression layer is itself a learning system.

The output is not "what most agents said." It is what the pattern of all agents said — and what that pattern has learned to mean.

Emergent Properties

• ·The brain wires itself. Frequently-solved problem types create fast, well-worn paths. Expertise is efficient routing, not superior raw intelligence.
• ·Fault tolerance by default. No Holon is critical. The system degrades gracefully rather than failing catastrophically.
• ·Specialization without siloing. Cross-domain synthesis happens naturally through shared pathways.
• ·Scale as intelligence, not cost. More Holons = richer network = more emergent capability. The marginal intelligence gain compounds.

The User Experience

The user sees nothing of this.

They see one interface: a single Holon. Behind it, 100,000 Holons have fired, communicated, voted, and converged. The answer arrives as though from a single mind.

Infrastructure

Three products. One architecture. One company: Intervals.

• ·Intervals.so — the OS. Auth, billing, memory, email infrastructure, scheduling.
• ·Holon — the agent product. Each instance is a Holon. The network is the brain.
• ·Magnets.sh — the spawner UI. Build and deploy Holons visually.

The moat is not the model. Models are commoditized. The moat is the trained network topology — learned pathway weights, reinforced neighbor relationships, a regression layer that has processed millions of queries. This cannot be copied. It can only be grown. Every query makes the brain smarter.

Open Questions

Conclusion

Holon is not a product improvement. It is an architectural bet: that the next leap in AI capability comes not from larger models but from societies of complete, communicating, self-organizing agents.

Human civilization did not advance by growing smarter individuals. It advanced by building institutions, markets, and communication networks that let ordinary agents produce extraordinary collective outcomes. Chaos — bounded, structured chaos — was always the engine.

References

Koestler, A. (1967). The Ghost in the Machine. Hutchinson.

Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.

Hebb, D.O. (1949). The Organization of Behavior. Wiley.

Hewitt, C. (1973). A Universal Modular Actor Formalism for Artificial Intelligence. IJCAI.

Simon, H.A. (1962). The Architecture of Complexity. Proceedings of the American Philosophical Society.

Koutsoupias, E. & Papadimitriou, C. (1999). Worst-case Equilibria. STACS.

Roughgarden, T. & Tardos, É. (2002). How Bad is Selfish Routing? Journal of the ACM.

Smith, A. (1776). The Wealth of Nations. W. Strahan and T. Cadell.