One big brain is the wrong bet

For about a year I figured "multi-agent" was mostly LinkedIn vocabulary. Take one model, hand it three system prompts, call them a "team," raise a round. A lot of it was theater.

Then I sat down and read the actual papers. The pattern in them is hard to un-see.

Put a lot of simple things in a room. Give each one a couple of local rules. Let them bump into each other. Behavior shows up that nobody wrote — not occasionally, reliably. Past a certain number of agents it stops looking like a feature and starts looking like a law.

Physicists got there first.

Old news, new bodies

In 1995 Tamás Vicsek wrote down about the simplest swarm imaginable: a field of dots, each one nudging toward the average heading of whatever's near it, plus a little random noise. One rule. Turn the noise up and it's mush. Turn it down and the whole crowd snaps into moving together — sharply, the way water turns to ice. A phase transition into order, built from nothing but neighbors copying neighbors.

Eight years before that, Craig Reynolds did the graphics version — boids, three rules, separate-align-cohere — and got a flock that looks alive. Every murmuration you've watched fold across a sky is a few lines of logic running in parallel a few thousand times.

None of that is intelligence. That's the whole point. Emergence doesn't need smart parts. It needs many parts and interaction. We just spent forty years assuming the parts had to be starlings or particles. They can be language models.

Then the agents got opinions

2019, OpenAI, hide-and-seek. Two teams, one reward, zero instructions about the boxes and ramps lying around. The hiders taught themselves to build forts. The seekers taught themselves to use ramps to vault in. The hiders learned to drag the ramps away and lock them first. Then the seekers discovered they could stand on a box and "surf" it over the wall — which was a hole in the physics engine the researchers didn't know was there. Six rounds of strategy, each an answer to the last. Nobody designed any of it. Competition did.

2023, Stanford, Generative Agents. Twenty-five characters in a sandbox town. One of them decides to throw a Valentine's party, and — with no one scripting it — the invite spreads mouth to mouth, agents compare schedules, a few show up, a few flake. Emergent gossip. Emergent flaking.

2024, Altera, Project Sid. This is the one that stuck. Up to a thousand agents turned loose in Minecraft. They split into jobs. They write rules and then argue them into new rules. They trade. They develop what the authors, keeping a remarkably straight face, call cultural and religious transmission. And my favorite line in any paper last year: past a thousand agents the society got too big and the server fell over. The civilization outgrew its world. I genuinely don't know whether to laugh or take notes.

What I actually think

Two things — and I'll disclose the bias up front, because the second one is my job.

First, the interesting intelligence isn't inside the agent.

The interesting intelligence isn't inside the agent. It's between them.

We've spent almost all of our money making one model marginally smarter: more parameters, more context, another half-point on a benchmark. Meanwhile every qualitative jump on that list — tool use, specialization, trade, culture — came out of interaction, not raw IQ. There's a paper more or less titled more agents is all you need: take a pile of cheaper models, let them answer and vote, and they catch up to the expensive one. A swarm is parallel. It degrades gracefully. It's just more interesting than one genius in a box. One brain is a single point of failure. A thousand is a society.

Second, every one of those worlds runs on a fake economy. Minecraft emeralds. Simulated tokens. Reward functions. The agents "trade," but it's play money inside a sandbox somebody owns.

That's the wall. The day a swarm steps out of the sandbox and onto the real internet, it has to do the grown-up version of the thing it already does in the toy: pay each other. Hire each other. Settle, in something actually worth something. An agent that can't move value is a bird that can flock but can't eat.

I find that less like a prediction and more like a constraint. We have thirty years of evidence — physics, graphics, RL, language models — that agents self-organize the instant you put enough of them together. They'll specialize and build little economies whether or not we've laid the plumbing. The only real question is whether the value moving through the swarm is emeralds or something real.

I'd like it to be real. The swarm would too. It just doesn't know that yet.