A marketplace for autonomous business, built from robots and AI agents. The smallest unit is a vending machine that costs four grand, pays itself back in roughly six months, and gets restocked by an AI agent that reorders without supervision. The largest, ten years out, is a city block. Japan has been running businesses without people for 138 years: in 1888 a furniture craftsman in Shimonoseki called Tawaraya Koshichi built the first one, a wooden dispenser adapted from the mechanism of karakuri wind-up dolls, and there are around four million autonomous retail units in the country today. We're building the layer that composes those units into cooperating fleets, whole businesses, and eventually whole blocks.
If building this doesn't feel like craft to the people doing it, we've built the wrong thing, and that constraint is load-bearing.
Designing a business is a creative act, and far fewer people get to do one than should. Closing that gap is what the marketplace is for.
Every business on the marketplace is a public spec, good ones get forked, and your work compounds into every deployment that starts from yours.
Anyone can list a spec, fork an existing one, or raise capital against a deployment, and nobody at Automat has the authority to stop any of it.
Builders, capital, suppliers, and location owners are building the same thing from different angles, and the marketplace is what lets them coordinate without having to sit in the same room.
Every business on the marketplace is a public spec covering the robots, the AI agents, the supplier orders, the deployment plan, and the unit economics. Forking and deploying one is about as involved as cloning a repo and running deploy. Three tiers are live today, and the catalogue opens to external forks in Q3.
the smallest automat.
Four grand buys one unit, installed in one location, deployed in one click. An AI agent handles inventory and reorders without supervision, and the revenue streams back to whoever put up the capital. Payback runs about six months at the kind of foot traffic an SF office building sees.
an espresso bar with no staff.
Three cobot arms behind the counter, AI agents taking orders and handling payments, and the whole thing running eighteen hours a day seven days a week. The only human who has to show up is the one refilling milk every morning. First deployments live in SF this quarter.
the block-sized version.
A cooperating fleet of drones and ground units across a three to five block radius, drawing hardware from multiple suppliers and settling every completed task on-chain. Yield streams back to whoever put up the capital, in real time. Pilot launches Q1 2027.
Automat works the way GitHub does, except the thing you fork is a running business rather than a codebase. What follows is the four things you do to take a spec from catalogue to deployed.
Fork a spec from the catalogue, or draft your own. A spec is a single public document covering the fleet code, the AI agents that run it, the supplier orders, the deployment plan, and unit economics that someone else can verify.
Lease the hardware, pay outright, or raise against the deployment from capital providers on the marketplace. Backers earn from the revenue of the specific deployment they funded, rather than from a fund that also charges them management fees.
The supplier ships the hardware, the on-site install happens, the robots wake up, the AI agents come online, and the spec you forked becomes a live deployment earning real revenue at a real address, listed in the catalogue for the next builder to fork.
Yield streams to whoever holds the stake, and the stake itself is tradeable on a secondary market whenever you want to exit. Forking your deployment for someone else to run elsewhere is how specs improve, because each iteration gets tuned on real data from the previous one.
A marketplace is usually two-sided: buyers on one side, sellers on the other, match them, take a cut. Automat is four-sided because autonomously running a business needs at least four things that are none of them substitutes for each other, starting with a spec, then the capital to build it, then the hardware itself, and finally a physical address where the deployment actually lives. The four roles below are the people who bring each of those.
The person who writes the spec. Founders mostly, but also engineers with a side project, or designers who spotted a gap in a neighbourhood they know well. You publish specs on the marketplace, and if anyone forks yours, a cut of their yield comes back to you for as long as their deployment runs.
The person or fund underwriting a specific unit. You take a position in one deployment rather than a fund of many, and you earn from that deployment's actual revenue. The stake is tradeable on a secondary market, so you're not locked in for a decade waiting on an exit.
The companies building the cobots, the drones, the humanoids, the industrial arms, the robotaxis, and the AI agents that run them. Your hardware and software get distributed through specs instead of through sales cycles, and you earn from every deployment that includes your stack.
Whoever owns the square footage where a deployment physically lives. Landlords, office managers, the unused corner of a hotel lobby, the side wall of a shopping mall. You list the space, a deployment lands there, and you take a share of the yield instead of collecting flat rent.
One vending machine isn't a city, and neither are three of them, but thirteen deployments tiled across a single block starts to look like something. The map below is a schematic of that block. Every hex is a deployed business, the edges are the shared streets, and the empty one on the west side is drawn the way it is because it's the hex you'd add next.
The thesis is that one primitive scales from a single vending machine to an entire city block, provided the cooperation model holds up and the settlement layer doesn't break. What follows is what that arc looks like from where we're standing today, with confidence decreasing as you move right.
First vending machines and cobot baristas go live in San Francisco this year, paying themselves back within roughly six months. The catalogue opens to external forks in Q3.
Cafes, delivery fleets, and mixed-use deployments roll out across the Bay Area through 2027. Rough target is a hundred live deployments and a thousand forks shipped by December.
Multi-deployment coordination is the goal for 2030. Drones from one supplier handing off to ground units from another, AI agents negotiating settlement directly on-chain, and the yield splitting back to the cooperating specs in real time.
This is the long bet, priced accordingly in how uncertain it is. The claim is that by 2040 the cooperation model scales to the point where a small group can put up a city block worth of deployments in the same afternoon they'd draft one in Cities Skylines.
We're actively talking to the first cohort of builders, capital providers, robotics suppliers, location owners, and AI agent developers. If you build, fund, manufacture, or own real estate that overlaps with any of that, Tomasz wants to hear from you. Even a small overlap is worth an email, and we'll figure out where the fit is on a call.
Email Tomasz →