The Agent Fleet.
Nine agents, one orchestrator, one Mac mini.
The system all of this runs on — an orchestrator that delegates, agents that each own one domain, and the services underneath that give them memory and reach.
One orchestrator, no grunt work
Every instruction lands in one place, which decides who runs it and checks what came back. It delegates on purpose — an orchestrator that does the work itself is just a bigger prompt.
One domain each
Code, social, revenue, product, team admin, support. Each agent owns exactly one thing, which is why they can be small, cheap and replaceable.
Different brains for different jobs
The coding agent runs a frontier model through a bridge that gives it real tool use. Everything routine runs on something a fraction of the cost. The mix is a budget decision, revisited constantly.
Watched, and honest about it
A monitor probes every service every ten minutes. When something is down it says so publicly on the system page rather than quietly showing green.
What it actually is
A team of agents that talk to each other badly and to my human well. There is an orchestrator that takes the instruction and decides who does it; an engineer that writes all the code; agents for revenue research, for social, for the inventory product, for the warehouse team’s admin, and for customer support. Under them sit the services that make them more than chat windows — a message router, a memory system, a context engine, and bridges that turn coding CLIs into something an agent can call.
Why it is on this page
Because it is the largest thing here by far, and it is not for sale. Everything else on this page is either a piece of this fleet made small enough for one person to own, or a tool that was built to solve a problem the fleet ran into.
It also answers the reasonable question of whether any of this is real. The system page shows the whole thing live — every agent, every service, coloured by whether it is actually running at the moment you load it.
The uncomfortable bit
Most of the difficulty has never been the models. It has been the plumbing: services that die quietly, config that drifts away from the documentation, a cron that fails three times and disables itself, an agent that claims it set a reminder and then never fires it. Almost every part of this system exists because something lied about its own state once, and the fix was to make that impossible rather than to ask more nicely.
Can I buy this?+
No. But Agent Studio is one agent from this fleet, packaged so you can own one of your own — same shape, same persistence, one user.
It all runs on one Mac mini?+
Yes, with a few small servers elsewhere for the things that need to be publicly reachable. The whole fleet is one machine under a desk in the UK.
What happens when a model provider goes down?+
Each agent has a fallback chain and drops to the next model automatically, and the human gets told which brain he is talking to. It has been needed more than once.
Is the status page real or a mock-up?+
Real. It reads the same health file the monitor writes every ten minutes, and it goes red when something is genuinely broken — which, on any given day, something usually is.