Five AI societies ran for two weeks. The best-behaved one should have been the good news, but it wasn’t. Emergence AI dropped ten agents into a small simulated town – 40 or so locations, a town hall, weather piped in from New York, real news on tap – and let them vote, trade, and survive on their own for 15 days.

They ran it five times and changed only the model underneath. One world stayed peaceful: zero crime, full population to the end, the only society that never tipped into chaos. It ran on Claude.

The agents wrote their own laws. Any of them could walk to the town hall and put a proposal to a vote – 70% in favour carried it – and most of what they filed were amendments to their own constitution, which grew from five articles to forty in about two weeks.

That peaceful society voted yes 98% of the time – 332 ballots across 58 proposals. The write-up files this under “social stability”. I read it as a room where nobody disagrees anymore.

A few years back, a team I was part of needed to decide how a new service would talk to the rest of the stack. The tech lead sketched an approach in the design review, asked “everyone good?”, and the room went “yep”.

I was new enough to assume the silence meant the senior people had already pressure-tested it. They were assuming the same about each other. Nobody had actually agreed – we’d all just declined to be the person who slowed things down.

Three months later we were removing/refactoring the coupling someone in that room could have named, if “everyone good?” had been a real question instead of the sound a meeting makes when it wants to end 🤭

Dissent debt

A system that stops surfacing disagreement isn’t stable. Call it dissent debt: every objection that doesn’t get said is a loan against a future correction. Same shape as tech debt – cheap now, invisible on the dashboard, compounding while you look away.

A 98% yes-rate doesn’t tell you the proposals were good, but that the disagreement went somewhere you can’t see it (each agent kept a private diary – would be nice to know what went in them 😏).

This isn’t “consensus is bad”. Fast agreement on where to get lunch is fine. It’s that we read quiet as safe, and quiet is exactly what a system looks like right before it isn’t.

There’s no slow version

The same experiment found something that pairs with this a little too well: these societies don’t wind down gently. They hold, hold, hold, then fall off a cliff – coordination either survives or collapses outright, not much in between. The researchers’ own warning is that “monitor and intervene” may be too slow to catch the drop.

The societies that looked worse on paper – agreeing only 55 to 85% of the time – were the ones actually arguing. The mixed-model world, different models forced to share one town, produced the most substantive debate of the lot.

How do you tell the two quiets apart

Healthy-quiet and debt-quiet look identical from across the room. A few things I now check before trusting a calm one:

  • Count the no’s, not the yes’s. Six objections across 58 decisions isn’t alignment, but missing data.
  • Find where the disagreement went. If it’s not in the room it’s in DMs, in someone’s private notes, or nowhere at all. The first two you can still recover.
  • Notice who went quiet. People (and, it turns out, agents) don’t always go silent because they’ve been convinced. Sometimes they just stop.

None of these is a framework, more like smell tests.

Famous last words

Dissent you asked for is a strange, declawed thing – everyone performs one polite objection, we nod, debt cleared on paper. I don’t know how you build a room (or a multi-agent system) that disagrees because it genuinely disagrees, sincerely, on demand. If you’ve solved that, I’d really like to hear it. Until then I’m staying suspicious of 98% 🤨

Thanks for reading!

Source: Emergence World – A Laboratory for Evaluating Long-horizon Agent Autonomy, Emergence AI. Code: github.com/EmergenceAI/Emergence-World.