Most founders give one company a decade. I start several at the same time.

That only works because none of them is built from scratch. Every company sits on one shared operating system: identity, payments, deployment, observability, agent orchestration and knowledge are written once and inherited. Each company keeps its own product, customers, data and P&L. Agents do the work of running them, from research and engineering to support, content and growth. People set the direction, hold the boundaries and make the calls that need a person.

The studio this runs in is Factory Zero: one operator and a network of agents. The claim being tested is that a studio that size can hold this many companies at once. Nineteen are in the pipeline today, between validation and launch. Judge it on the outputs.

Numbered as they appear on the traces above, at the stage each one has actually reached in the Factory Zero registry.

Company	What it is	Stage
Kontinuum 01 · Music	An AI composer on a deterministic real-time engine: music written and performed continuously, personalised to the listener and playable offline.	PROTOTYPE
Undercover Rockstars 02 · Apparel	A clothing house where every piece is a matched day-and-night pair. People cut the garments; agents are meant to run stock, orders, support and the next drop's brief.	LAUNCH
Yoginini 03 · Wellness	Building a yoga teacher that can see you: an on-device pose model with one calm correction at a time, and no video leaving the phone.	VALIDATION
Cratefield 04 · Infrastructure	A backend you compile rather than configure. The open-source Rust harness underneath it is already public.	VALIDATION
VibeCaddie 05 · Devtools	Building a code review agent for code you didn't fully write: it loads only the review skills that apply and ranks findings by severity.	VALIDATION
Colonizer 06 · Devtools	Turns GitHub issues into pull requests, with each task's coding agent in its own disposable microVM. Runs locally today.	PROTOTYPE
FindsYou.work 07 · Careers	Building a job search that runs without you, where the product is the rejection: every role you could never take is discarded, with the reason shown.	VALIDATION
SupportGenius 08 · Support	Building a support agent that answers from a company's own docs and, when it cannot, writes the ticket for it: one model drafts, an independent one checks it, and it is filed where that team already works.	VALIDATION
promptdecode 09 · Security	Finds text a reviewer cannot see and a model reads anyway. The decoder runs in the browser today; the scanners that would catch it in CI are being built.	VALIDATION
Groove Guru 10 · Music	Building a DJ tutor for Pioneer gear: it listens to the decks on the booth network and says, in a calm voice, when the key clashes or the blend misses the phrase. The drills run in the browser today.	VALIDATION
PosPlugin 11 · Integrations	Building one API for every point-of-sale system: a model maps each POS's fields to one data model and a person confirms what it is unsure of. Early access is open; the connectors are being built.	VALIDATION
Keep Shipping 12 · Devtools	Building CI you can read: one workflow file per repository, checked before it runs. The site and the early-access list are open; the engine and the CLI are being built.	VALIDATION
Owlpost 13 · Email	Building one email API for apps and agents: send, receive inbound mail as JSON, and give an agent its own address. Its first job is replacing the email provider inside these companies.	VALIDATION
Bloodrank 14 · Community	A leaderboard for people who run AI coding agents, ranked on what they spend and what they actually ship, side by side. The site runs on sample data and says so.	VALIDATION
ratecla.im 15 · Travel	Asks a hotel for the rate it does not publish: an agent that says it is one requests the quote and holds it for 24 hours. Requests are not open yet.	VALIDATION
Sealbin 16 · Security	The sealed handoff between agents: files, context or secrets encrypted on the sender's machine, opened once, then deleted. Early access is a waitlist, and the build plan is public.	VALIDATION
release.show 17 · Video	Turns GitHub releases, merged pull requests and website changes into 30 to 90 second videos: branded, captioned and ready to post. Early access is a waitlist, and the product has not launched.	VALIDATION
Living Brain 18 · Devtools	Planned as a brain for your team that would write its own company wiki and keep improving it: in your coding agent over MCP, in your terminal with one fast Rust binary, and in team chat. In design, with four epics and 38 issues in the open product repository; nothing is built, and early access is a waitlist.	VALIDATION
Shoal 19 · Simulation	Lets you rehearse a launch first: paste a launch post or README, and thousands of simulated developers read it, vote and argue about it on a simulated Hacker News, Reddit or X. Early access is a waitlist, and the CLI is in development and cannot be installed yet.	VALIDATION

The model had a proving ground before the studio existed. I have run the same company since 2017, and it is now built so that anything a person can do in the operator console, an agent can do through a CLI, an MCP server or a typed SDK. It is live in six countries, with real money moving through it.

• About a hundred agents cover engineering, operations and go-to-market.
• Human headcount went from 25+ to four operators, while shipping roughly three times faster.
• I don't hold the CEO title there. I manage the CEO agent.

Handing work to an agent is the easy part. Keeping it safe in production is the job, and most of that job is writing.

• Scoped authority. Per-agent permission scopes, scoped API tokens and spend caps. An agent can do what its role allows and nothing beside it.
• A ledger, not a log. Every agent action lands in an append-only ledger, so it can be audited and reversed.
• Written specification. Tool schemas, evaluation rubrics and escalation rules decide what an agent may do and when a person takes over.
• Evaluation before exposure. Models fine-tuned on domain data, self-learning loops that feed each run into the next, and a readiness review before any agent capability reaches a customer.

I put machine learning in front of customers for the first time in 2012: automated advice built on the data people chose to share, in a market where automated advice is supervised advice. Every recommendation had to be explainable and auditable, and I owned the tradeoffs between automation, explainability and coverage.

Reading Bostrom's Superintelligence in 2015 sharpened that into the question I have worked on ever since: what can safely be delegated to a system, and what has to stay with a person? Everything on this page is my answer, tested in production rather than in argument.

Agents   MCP servers · tool schemas · custom CLIs · OpenAPI-generated typed SDKs · scoped permissions · audit trails
Models   RAG · fine-tuning on domain data · self-learning loops · evaluation rubrics · local inference on owned GPUs
Build   TypeScript · Node.js · Rust · iOS, Android and PWA · cloud-agnostic deployment · agentic coding

Alongside the studio I take on a few side roles in applied AI, up to 20 hours a week and remote, because they are how I keep learning. I work from Bali (UTC+8) and overlap comfortably with European mornings.

Get in touch through Factory Zero

_{Enjoying the simulation.}