Swarm Intelligence (AI)

From one model to a swarm

The dominant framing in AI through the early part of this period was a single, increasingly powerful general model handling every task. A different architecture emerged quickly around the launch of open agent frameworks: rather than one generalist model, a swarm of specialized agents that collaborate, negotiate, and delegate to each other.

Peter Steinberger explains the shift by analogy to human civilization, in conversation with Garry Tan: "if you look at one human being, what can one human being actually achieve? Do you think one human being could make an iPhone or go to space? One human being would probably just not even be able to find food. But as a group we specialize, and as a larger society we specialize even more."¹ No single person can assemble a smartphone from raw materials; it is a product of civilization's division of labor, not of any one person's individual capability, and the same logic scales to AI: a swarm of specialized agents, each with deep competence in a narrow domain, can accomplish things a single generalist model cannot.

What it looks like in practice

In a working implementation, one agent manages a person's private life, another manages work, another manages a calendar, sharing context where needed but otherwise operating inside its own domain. Agents also negotiate directly with each other: a personal agent might reach out to a restaurant's own booking agent, and if the restaurant has no agent, hire a human to place the call instead. Community-built extensions of open agent frameworks already demonstrate this kind of bot-to-bot negotiation working in practice.¹

Steinberger raises a subtler version of the same question: even a single, generalist foundation model might be better understood as behaving like a specialist for most individual tasks, so the interesting design question is not only whether to build a swarm of separately fine-tuned models, but whether one underlying model can behave as a genuine specialist when instructed narrowly enough. The swarm insight, on this reading, is primarily about coordination architecture rather than strictly about the number of distinct underlying models.

Why it matters

The framing shifts the central product question from how capable a single model is to how well an entire swarm coordinates, and the ceiling on a swarm's capability comes down to three things: the quality of specialization inside each individual agent, the communication and negotiation protocol between agents, and trust, meaning how one agent knows it can rely on another it is hiring or negotiating with. For builders, the implication is that a specialist agent excelling in one narrow domain, booking, coding, or finance, is a more durable position than attempting to build a single generalist competing directly with frontier labs, because a swarm architecture creates ongoing demand for exactly that kind of specialist rather than replacing it.

Open questions

How much of a swarm's theoretical capability gain gets eaten by coordination overhead is an old organizational-design problem restated in a new substrate, the same reason large companies are not proportionally more productive than small ones. Trust between agents transacting with each other needs a mechanism, cryptographic, reputation-based, or both, and remains largely unsolved infrastructure. Who owns a given swarm is also unclear once its component agents are hosted across different providers, which fractures any assumption that an agent's data and behavior stay under one party's control. And a swarm of interacting agents can produce emergent behavior that no single agent in it produces on its own, which is exactly what makes the architecture powerful and is also its most direct coordination-failure risk.