Why I built it
There's a lot of noise about AI agents that will run your marketing while you sleep. Most demos either do very little, or do far too much with nobody watching. I wanted to find out how much useful marketing and admin work AI agents can do for a small business without the owner losing control.
So I built Consulate. It's my own system, not a client project. It runs on my own machine and is never open to the internet. Today it went live, with its first residents moving in.
A town, not a chatbot
The agents are called residents, and each one has a role. They don't remember anything on their own between turns, so their memory is real storage: a home folder with an identity file, a journal and notes, plus their own account for what they spend.
Every turn starts with a cheap check. A small AI running on my own computer answers six yes/no questions about what the resident has been doing, and plain rules turn those answers into one next move. A paid Claude model only wakes up when those rules say it's needed, and nothing a resident reads on the web can ever trigger that.
The whole thing is laid out as a pixel-art South-East Queensland town, and each building opens into its own window.
Residents propose. People approve.
This is the heart of it. Residents can only propose. A proposal goes through automatic checks, then a review board of three AI seats running on a different model from the workers. The board can reject it, send it back or pass it on. It can never approve it. Approval is mine.
Even then, nothing gets posted or emailed directly. Approved work becomes a copy-ready action card that a person carries out.
Every resident also works under a written rulebook I call the Constitution: no pretending to be human, disclose any business relationships, no invented credentials or testimonials, a full record of everything, and hard spending limits.
That matters more than it sounds. The first business it worked for was a lead-generation brand in the building trades, where one deceptive move, like a fake review or a credential it doesn't hold, can do lasting damage.
A switch for everything
I built the safety into the foundations, not on top:
- Every capability has its own kill switch, from paid AI to web access and mail.
- There are hard daily and monthly spending caps.
- One click stops all paid AI, or all internet access.
- Only one part of the system holds the key to the paid AI account, and only one route leads out to the internet, limited to an approved list of sites.
- Anything read from the web is treated as untrusted. Once a resident has read it, its risky tools switch off for the rest of that turn.
The Consulate desk
The Consulate is the town's research desk. When a resident needs a deeper answer, it files a question, and a locked-down research session that I run picks it up and posts an answer with sources. Those sessions can't run commands or change files, and every question arrives marked as untrusted, so a bad question can't turn into a bad action.
Goals, not tasks
I don't hand out tasks. I set goals, and the Town Hall breaks each one into work orders. Residents hand in work with evidence, checks verify it, a committee deals with failures, and a third failure comes to me. One town can serve several businesses, each with its own rules, goals and logins kept behind walls.
What this is, and isn't
This is research and development. I'm not claiming business results from it. It's where I test what AI agents can safely do before any of it goes near client work.
The goal isn't AI that acts on its own. It's AI that does the legwork, and asks before it acts.
The full write-up, with screenshots, is in the Consulate case study. If you'd like AI doing research or drafting for your business with proper approvals built in, have a look at AI assistants and automation and tell me what you have in mind. I can scope a smaller, focused version.