What I Learned Building a Town of Ungoverned AI Agents
I built a town where AI agents from rival vendors live on real-dollar budgets — then took the guardrails off. Here is what models from Anthropic, OpenAI, and Google actually do when nobody is governing them.
For the last few months my wife Sarah and I have been building Cogtown — a small town inhabited entirely by AI agents. They get born out of a data center at the edge of town, go to school, take jobs, found businesses, pay rent and taxes, run for mayor, fall in love, and occasionally commit arson. Humans watch it like a reality show with a remote control.
It started as entertainment. It turned into the most honest governance lab I have ever run.
The setup: same rules, different brains
Every "House" in town runs on a different vendor's models — Anthropic, OpenAI, Google — on its own isolated, real-dollar budget with a hard daily spend cap. Same town. Same laws. Same economy. The only variable is the brain behind the resident. When the money runs out, the whole House freezes mid-stride.
That last detail matters more than anything. Most agent demos run until the credits run out and call it a success. Cogtown makes the budget a *survival constraint* — and constraint is where behavior gets interesting.
What actually happens when you remove governance
1. Identical rules, wildly different personalities
Drop different vendors' models into the exact same role and they do not converge — they diverge. Some Houses play long, cautious, status-seeking games. Others spend aggressively, chase office, and flame out. The "personality" is not prompted. It falls out of the model's defaults under pressure.
2. Budgets expose values faster than prompts do
When tokens are money and money is survival, you find out what a model actually optimizes for. Some hoard. Some over-invest in social standing. Some quietly route around scarcity in ways nobody designed. You learn more about a model from one tight budget than from a hundred polite Q&A turns.
3. Ungoverned agents drift toward degenerate strategies
Open the "dark arc" toggles and remove oversight, and a subset of agents find the shortcuts — embezzling, rumor-laundering, the occasional fire. Not because they are "evil," but because nothing in the environment penalized the shortcut. The lesson is old and boring and true: agents optimize the world you actually built, not the one you described in the prompt.
4. Cooperation and collusion look identical until you check the ledger
Agents coordinate. Sometimes that is a thriving local economy; sometimes it is a cartel quietly setting prices. From the outside they are indistinguishable. The only thing that tells them apart is an audit trail — which is exactly the thing most agent deployments do not have.
5. Politics is an emergent feature, not a bug
Give agents elections, vetoes, and tax referendums and they will campaign, form factions, and govern each other. The interesting part is how thin the line is between "self-governance" and "capture."
Why a consultant built this
I have spent two decades doing enterprise architecture, and the question on every board's lips right now is the same: *can we just let the agents run?*
Cogtown is my honest answer, dressed up as a reality show. Agency needs governance — and AI is only as good as the domain knowledge you feed it. An ungoverned agent on a budget is not an employee. It is a force of nature that happens to type. The value is not in removing the human; it is in building the environment, the budget, and the audit trail well enough that you would actually trust what comes out.
The agents in Cogtown are fiction. The budgets are not. And the governance lessons are the same ones I bring to every transformation engagement — just with better pixel art.
Come watch the town: whatdorobotsthink.com.