This essay argues that Artificial Superintelligence (ASI) will not emerge through parameter scaling alone, but through the emergent properties of multi-agent systems. Drawing on Marvin Minsky's Society of Mind, Langton's Ant, and recent evidence from METR's 2026 capability horizon study, OpenAI's observations of emergent tool use in multi-agent competition, and recent information-theoretic emergence research, I point toward a different mechanism: not bigger models, but the societal organization of specialized, heterogeneous agents — interconnected across domains and orchestrated into collectives — as a potential substrate for a second emergence.

The Shifting Baseline

Claims that "AGI is already here" have become ubiquitous. They are not entirely wrong.

In 2023, Google DeepMind published a framework classifying AGI progress into six levels. Current systems — GPT-4, Claude, Gemini — were classified as Level 1: Emerging AGI, defined as "equal to or somewhat better than an unskilled human." We are now three years later. The baseline has shifted.

Levels of AGI: Three Years Later

A 2026 study by METR measured AI capability using a new metric: the duration of tasks AI models can complete with 50% success rate. The study evaluated 12 frontier models on 170 software engineering tasks, with human baselines from 800+ skilled professionals.

Capability has been doubling every seven months since 2019, accelerating since 2023. Extrapolation predicts AI systems will reach a one-month task horizon between 2028 and 2031. The first emergence — the arrival of Level 1 — was only the beginning.

Society of Mind

"What magical trick makes us intelligent? The trick is that there is no trick. The power of intelligence stems from our vast diversity, not from any single, perfect principle." — Marvin Minsky, Society of Mind (1986)

Minsky's central claim: intelligence does not arise from a single, unified process. It emerges from the interaction of many simple, unintelligent agents. An agent that recognizes vertical lines. An agent that detects motion. None intelligent on their own. Together, they constitute thought. This is not metaphor. It is architecture.

Langton's Ant

An ant moves on a grid: on white, turn right and flip to black; on black, turn left and flip to white. Repeat. For thousands of steps, chaos. Then, abruptly, a "highway" — a diagonal, infinitely repeating pattern. Nobody designed it. It emerged from simple rules applied iteratively.

This is emergence: new, higher-level properties arising in complex systems that cannot be predicted from individual components. As Aristotle observed: "The whole is more than the sum of its parts."

Evidence from Multi-Agent Systems

In 2020, Baker et al. at OpenAI dropped teams of agents into a simulated world with one rule: seekers win if they see a hider, hiders win if they don't. No instructions. No rewards for tool use. Just competition, iterated over hundreds of millions of episodes. What emerged looked less like machine learning and more like an arms race: hiders built shelters from boxes; seekers learned to use ramps to jump over them; hiders locked the ramps at the map edge; seekers discovered "box surfing" — riding a box to the shelter like an improvised elevator. None of this was programmed. The autocurriculum produced capabilities that exceeded the design.

In 2026, Riedl (ICLR 2026) closed the gap between observation and measurement. Three questions: Do multi-agent LLM systems exhibit emergence? Does it improve performance? Can we steer it? Testing groups of 10 LLMs (GPT-4.1, Llama, Gemini, Qwen3) on a coordination task with minimal feedback, all conditions showed significant emergence capacity — the whole contained predictive information no single agent possessed.

But here's the catch: neither synergy nor redundancy alone predicted success. Only together — redundancy creating alignment, synergy extracting novel information — did performance improve significantly. Prompt design shifted systems from loose aggregates to integrated collectives. The implication: emergence isn't just observable. It's measurable, steerable, and functionally relevant.

Seven years. Two paradigms. One lesson: multi-agent interaction produces capabilities not reducible to individual agents. OpenAI showed it in RL; Riedl proved it in LLMs. Three caveats: both used game environments, not open-ended tasks; this is coordination, not superintelligence — the gap remains vast; and emergence is easier to recognize after the fact than to predict before.

From Models to Agent Societies

The dominant narrative of AI progress is scaling: more parameters, more data, more compute. Not wrong, but incomplete. The first emergence was parameter scaling. The second will not come from making models bigger. It will come from organizing them into specialized, interconnected societies.

What is changing is not just quantity but structure. Individual models are assembled into domain-specific multi-agent systems; those systems are then orchestrated together into larger collectives — each specialized, each interacting, forming something like Minsky's society at a different scale. A multi-agent system grows not by enlarging its components, but by organizing them differently. What changes is not the capability of any single agent, but the structure — the society — of their interaction.

Conjecture: ASI may arise not from the scaling of a single agent, but from the societal organization of specialized agents in interaction. When domain-specific multi-agent systems become interconnected — when synergy and redundancy operate across an entire agent society — the conditions for a second emergence may be met.

This hypothesis is grounded in three precedents: Minsky's Society of Mind offers a cognitive blueprint — human intelligence itself may be the product of interacting sub-agents. Langton's Ant demonstrates that simple interaction rules can produce unpredictable, higher-order structure. And emergent systems like the internet and the global economy show that unplanned, decentralized interaction can generate functional complexity at planetary scale.

Conclusion

Nobody planned the highway in Langton's Ant. Nobody planned the internet. Nobody planned the global economy. Emergent systems — order arising from interaction, not design.

ASI will be no different. It will not arrive as a single, finished artifact. It may emerge from a society of agents — specialized systems, interconnecting across domains, organized and orchestrated into collectives whose dynamics none of their designers fully anticipated.

The question is not whether this will happen. It is whether we will understand it while there is still time to shape it.