How does one person run platform-scale infrastructure?
Seventeen newspapers, a SaaS platform, an agency, and a toolchain — one operator. The stack decisions and operating rules that make it hold.
By refusing, at every layer, to own problems that someone else operates better at scale — and by treating AI agents as staff, with the management structure staff require. That's the whole answer. I run a network of 17 local newspapers, the multi-tenant platform they sit on, a client agency, and the internal tooling that connects it all, as one person. Here's what that actually rests on.
Rule one: own zero servers you don't have to
Everything customer-facing runs serverless — Next.js on Vercel, Firebase underneath. Not because it's fashionable, but because of what I don't do at 2 a.m.: patch boxes, rotate certs, babysit a database failover, scale for a traffic spike when a story runs. I ran infrastructure the hard way for years — in the nineties I founded and operated an ISP, which is about as far into "you own every problem" as this industry gets. The lesson from those racks: every component you operate yourself is a pager you wear. A solo operator has one pager and it must be reserved for the things only they can fix.
The same logic picked the architecture. Multi-tenancy means 17 papers are one deployment with 17 configurations, not 17 systems — fix once, fixed everywhere; paper #18 is a config file. Boring, proven services beat clever self-hosted ones every time the two compete, because boring is a synonym for "someone else already found the failure modes."
Rule two: agents are staff, so manage them like staff
The genuinely new part. The daily work of a 17-paper newsroom — sourcing, drafting, linting, publishing — runs through AI pipelines. Client and internal dev work runs through coding agents executing scoped work orders. What makes that hold isn't the models; it's the management structure around them, which I've written up in detail in the pipeline dispatch: a planner-reviewer at the top that never executes, cheap executors that never set scope, written handoffs at every boundary, review gates before anything ships, and a human approval gate on everything irreversible or public.
The mental shift that matters: stop asking "what can AI do for me?" and start asking "what would I delegate if I had staff, and what would I need to see before trusting the work?" The second question forces you to write real specs and real acceptance checks — which is precisely the discipline that makes delegation to machines work. People who can't delegate to agents usually couldn't delegate to people either. The bottleneck was never the staff.
Rule three: the operator does only operator work
Strip away what's delegated to services and agents, and what's left is the actual job: deciding what gets built, setting the standards work must meet, reviewing what ships, and answering for all of it. Judgment, taste, and accountability. Everything else is either a managed service's problem or an agent's task.
That's also the honest limit of the model. One person can run platform-scale systems; one person cannot be a platform-scale institution. I don't pretend the papers have a newsroom of twelve — the architecture is built so that the humans who matter (editorial accountability above all) sit at exactly the gates where humans are irreplaceable, and nowhere else. Solo-scale isn't about removing people from the loop. It's about spending the only person you've got exclusively where a person is the point.
Could you do this?
More people can than currently believe it — that's the window this whole platform shift has opened, and thirty years of these shifts say the window eventually prices in. But the prerequisites are real, and they're not the ones usually advertised. You need enough engineering depth to judge work you didn't do line-by-line, because review is now your primary technical act. You need the discipline to write specs, run gates, and keep handoffs written when nobody is watching you skip them. And you need to genuinely accept that your leverage comes from the boring choices — the managed service, the one deployment, the config file — because the exciting choices all come with pagers.
One operator, platform scale. It's not a trick. It's subtraction, done relentlessly, until the only work left is the work that needed you.