Working solo with AI
Solo AI developer: what it actually looks like
Can one person with a fleet of AI agents really run several products at once? It's a fair question, and I'd have doubted it too.
So here's the honest version, with me as the worked example. I'm Jack Stovell, I'm in London, and I run a one-person software studio called Adapt Progress Evolve Limited (company 10047328, incorporated 7 March 2016). If you'd rather see it than read about it, there's the 60-second tour.
What a solo AI developer actually is
A solo AI developer is one human who builds and runs software products with AI agents doing a big chunk of the actual work. That's the definition. No magic in it.
But the phrase hides a gap between what people picture and what it looks like in practice. People picture a person typing a sentence and a finished product falling out. What actually happens is closer to running a very small, very fast, slightly overconfident team, except the team is software and you're the only one who can say yes.
That last part matters. One person plus twenty-six agents isn't the same as one person working alone, and it isn't the same as a company either. The agents can research, build, audit and check. They can't decide what's allowed to go public. That stays with me, and I think it's the whole trick.
How my studio actually runs
The studio runs on a schedule, and the schedule is the thing that makes it feel like a studio rather than a pile of chatbots.
At 06:10 a brain brief rewrites what the studio knows for the day's agents. Think of it as a fresh page of context, so nobody starts working from stale assumptions. At 07:30 there's a morning brief. The voice QA looks back over yesterday's Jarvis conversations to see what went well and what didn't. Scheduled posts publish at 08:15.
A planner sweeps three times a day and proposes work, and that work goes into a queue for approval. It doesn't just start. SEO, AI-visibility and security audits sweep the live sites. And every ten minutes a deploy sweep pushes anything that's been merged.
Is the approval queue a bottleneck? Absolutely, and I'd like to keep it that way. If everything flowed through without a human yes, I'd have no idea what my own studio was saying in public.
What the agents do and what I do
The rule is simple to say. Agents research, build, audit and check. Nothing ships without a human yes, and the human is me.
The fleet is 26 Claude agents, and they've got roles, like proper job titles. Jarvis is the voice-first assistant that coordinates everything. There's memory and a planner, plus builders, SEO, AI-visibility checks, a writer and an editor, design QA, security, infrastructure, docs, research, a news desk, and then an inbox, a calendar and the morning brief.
Reading that list back, it sounds like a much bigger operation than one person in London. And that's sort of the point. Each agent does one narrow job well enough that I can trust its output as a starting point.
What do I do? I decide. I read the proposals the planner puts in the queue and say yes, no or not yet. I look at what the builders have made before it goes anywhere near the public. I pick the direction. The agents bring me options and evidence; I bring the judgement and, frankly, the taste.
It's less "I built this" and more "I approved this, and I'm accountable for it". More like editing than writing.
What it has shipped so far
There are 17 projects on jstov.uk/work. Fourteen are live: ScriptGrain, ReviewNudge, Hilyt, visibilitytoai, Launch Video Maker, Forematter, Assets for three.js, Marketframe, Paws of London, How to get AI to…, What is GEO?, GLA Spend Explorer, Cortex Map and Jarvising. One is in development, which is Jarvis. Two are paused: UKMWI and Football Scout Engine.
Paused is an honest word. Not every idea earns the attention it wants, and having agents doesn't change that. It just makes it cheaper to find out.
Each project has its own deep dive at jstov.uk/work, if you want the detail on any of them.
For a sense of pace, here are the last 30 days of commits, measured on 29 September 2026: ScriptGrain had 229, Forematter 128, Launch Video Maker 103 and Paws of London 65. Commits aren't a perfect measure, and a big number doesn't mean a good product. But it does show the fleet moving several products forward at the same time, which is the bit people doubt.
The stack
Nothing exotic here, which I quite like.
Claude runs the agents and the voice assistant. The sites are React and Vite, or Astro, depending on what the site needs. Hosting and edge logic sit on Cloudflare Pages and Workers. Supabase and Postgres handle the data. Blender makes the 3D assets. And a VPS runs the agents, so the fleet keeps working whether or not my laptop is open.
That's the lot. I'd rather have a stack I understand end to end than a clever one I can't debug.
The Jarvis architecture is public, if you want to see how the coordination side hangs together: github.com/StovBuilds/jarvis-architecture. There's also a weekly series on it, Jarvis: the architecture series.
What goes wrong
Plenty. Let's not skip this bit.
Agents invent confident detail if you let them. That's the big one. They don't hedge; they just sound sure. My own product ScriptGrain wrote copy for this site in my voice, and a fact-check still had to repair 12 lines it had embellished. Twelve, from a tool I built, writing about me. So agents need checking, always, no matter how much you like them.
Jarvis gets things wrong too. I've written about it in What Jarvis actually gets wrong (and why I'm not fixing it yet), because some fixes are deliberately deferred. Not everything that's broken is worth fixing today, and pretending otherwise would be dishonest.
Then there's memory. If it isn't cleaned up, old completed tasks resurface as new proposals. You'll be sitting there thinking, didn't we already do that? Yes, we did. The planner just doesn't know it's finished.
None of this is a reason not to work this way. It's a reason to build the checks in from the start, and to be relaxed about admitting the stuff that's still rough.
How to start if you want to work this way
Don't build a fleet. I mean it. Start with one agent doing one repeatable job, something dull that you'd otherwise do by hand every week. Get that working and trustworthy before you add anything else.
Then write the rules down, and put them somewhere every agent reads. Not in your head, not in a chat that scrolls away. Written down, in one place, so that the tenth agent behaves like the first.
Log every promise. If an agent says it'll do something, that goes in a log, so the work survives when a session dies. Sessions do die, and if the only record of a promise was inside a crashed conversation, it's gone.
Put a human approval gate on anything public. This is the one I'd never bend on. Research and drafts can run freely, but the moment something is going to be seen by other people, a person says yes.
And fact-check AI copy line by line against a fact sheet. Not a skim. Line by line. Remember those 12 lines.
That's really it. One agent, written rules, a log, an approval gate, a fact-check. Do those and you're a solo AI developer, or at least you're starting to be.
Is it what people imagine? Not quite. It's quieter, more about checking than typing, and a lot more like editing than magic. But it's real, and it's running right now.
FAQ
What is a solo AI developer?
A solo AI developer is one person who builds and runs software products with AI agents doing a large share of the work. The agents research, build, audit and check; the person decides what ships.
How many agents does Jack Stovell run?
Jack Stovell runs 26 Claude agents in his one-person studio, Adapt Progress Evolve. Roles include a voice-first assistant called Jarvis that coordinates, a planner, builders, SEO, security, design QA, a writer and an editor.
Can AI agents ship things without a human?
Not in this studio. The rule is that agents research, build, audit and check, and nothing ships without a human yes. Work the planner proposes waits in an approval queue.
What goes wrong when you work with AI agents?
Agents invent confident detail if you let them, so AI copy needs a line-by-line fact-check. Uncleaned memory can also make finished tasks come back as new proposals.
How should I start working as a solo AI developer?
Start with one agent doing one repeatable job. Write the rules down where every agent reads them, log every promise so work survives a session dying, and put a human approval gate on anything public.
Last checked: