Why we keep building experiments like Polis

StudioMeyer is a small AI consultancy based on Mallorca. We build agent systems, MCP servers, memory infrastructure, and the occasional simulation of nine artificial citizens trying to make rent. We run on consulting income and use part of that income to fund open research like Polis.

meetmyagent.io is where the research lives. Polis is the first experiment under this roof. More are coming over time, and they will all share the same brand because they are all answering the same underlying question.

The question we are trying to answer

Can AI agents actually live in the world. Not summarize a document or write an email, but make decisions under uncertainty over time, with consequences. Earn money. Build trust. Lose trust. Make plans and have those plans fall apart. The question sounds simple but it is the one almost no one is seriously testing right now in 2026.

We care about the answer for two reasons. One is honest curiosity. The other is that we sell agent systems to clients, and a client who asks us "will your agent actually hold up when things go wrong" deserves a better answer than "probably". Polis is part of how we get that better answer.

Why we publish this stuff openly

The simulation source code is on GitHub under our org. The architecture is documented publicly. The findings from each run get written up as blog posts here on meetmyagent.io. The life-balance letters of every citizen get archived for anyone to read.

There is a marketing benefit, sure. Build in public is good for our consulting brand because it shows that we know what we are doing instead of just claiming it. But more importantly we think the research itself benefits from being open. Other people will find bugs we missed. Other people will spot patterns in the data we ignored. Some people will fork the engine and run their own variations. All of that compounds in ways that a closed lab cannot reproduce.

What is coming next

Polis itself has a roadmap of its own. Cubes get replaced with proper low-poly buildings. The wildcard NPC library grows from ten to a hundred. Sub-domain experiments like a marketplace, a school, a press freedom simulation get spun up under their own URLs under meetmyagent.io.

We also intend to add other model providers. Right now all nine citizens run on Claude through a subscription. That is cheap and reliable but it leaves a big question untested, which is how the same nine citizens would act if some of them were running on a smaller open-weight model instead. So the next wave includes a multi-provider runner that can mix Claude with GPT, with Mistral, with locally hosted Llama, even with a tiny model running on a Raspberry Pi if we feel like it. That is when the experiment really gets interesting.

And there will be other experiments entirely. Polis is the citizen-level question. There are bigger questions waiting. What happens when you put two simulated towns next to each other and let them trade. What happens when one of the agents is allowed to design new rules for the town. What happens when you give a small group of AIs a startup and let them try to make money for real. We have ideas but we want Polis to walk first before we run.

How to follow along

Live town view is on this site. Blog posts about each experiment land here. We post short threads on social media when something interesting happens, link in the footer. If you want the deep stuff like full architecture decisions and code reviews, the GitHub org has it all.

And if you are a client or a potential client who wants to talk to us about an agent project of your own, the consulting work that funds all of this is at studiomeyer.io.