What Is a Dark Factory?
A dark factory is a factory that runs with the lights off. No people on the floor, so no need for lighting. The machines run the line. Applied to AI, it means a system of agents that keeps working while you are not watching, on hardware you control.
Where does the term come from?
Manufacturing. The industry phrase is "lights-out manufacturing": a plant so automated it can run in the dark, because nothing on the floor needs to see. One well-known example is FANUC, the Japanese robotics maker, which has reportedly run robot-assembly lines lights-out since 2001. Machines building other machines, for weeks at a stretch, with humans checking in on the results rather than standing at the line.
The term is not a metaphor for mystery or secrecy. It is a description of an operating state: the system runs without supervision, and the output is what you inspect.
What does a dark factory mean in AI?
The same operating state, applied to AI agents. A dark factory in AI is a set of agents with an orchestrator (the agent that assigns work to the others), a task board, and shared memory, running on your own hardware. Tasks go on the board. Agents pick them up, do the work, and post results for review. You are not in the loop keystroke by keystroke. You assign, then you inspect.
That is the difference between this and a chatbot. A chat session is you standing at the line: every step needs your next message, and when you close the tab the work stops and the context evaporates. A dark factory keeps state in files and a memory store you own, so the work survives the session, the reboot, and the model you happen to be using this month.
What a dark factory is not (and how it compares)
| Approach | Who owns it | Where it runs | What persists |
|---|---|---|---|
| Chatbot session | The provider | Their cloud | Nothing after the tab closes |
| Hosted agent SaaS | The vendor | Their cloud | Your data, on their terms |
| Agent framework code | You, if you build it | Wherever you deploy | What you wire up yourself |
| Dark factory | You | Your hardware | Memory, skills, task history |
The honest row in that table is the third one. You can build all of this from a framework. The difference is that a framework hands you parts, and most people stall at the wiring: persistence, orchestration, cost control. A fuller version of this comparison, including what Dark Factory is not, is on the compare page.
Why run agents this way?
- Ownership. The system is plain files on your machine. Change providers, change models, keep everything: memory, skills, workflows.
- Persistence. Agents that remember yesterday are useful in a way that stateless chat (a session that starts from zero every time) is not. The memory is the asset, and whether you buy or build it is its own guide.
- Cost control. Unattended agents burn tokens, the units AI usage is billed in. Owning the stack means you can route work to local models (AI running on your own machine, with no per-request bill) and watch the spend. We published the numbers from our own meter, and the build log records what that looked like in practice, including the mistakes.
How do you build one?
Incrementally. Nobody should start with a fleet of agents.
New to all of this? Start with the free primer. It covers the thinking: why chat has a ceiling, and how assigning work to agents differs from asking questions.
Ready to build? Start with one assistant that has persistent memory and a daily routine. That is Level 1: about 10 hours, no coding, your AI agent does the building. Then scale that assistant into a factory: an orchestrator, a worker roster, a task board, and a shared memory hub; how that orchestration works without LangGraph or CrewAI is its own guide. That is Level 2: about 13 to 14 hours, running unattended on hardware you control.
This site is itself run as a dark factory. Start with the free primer, or build your own with Level 1. The February 2026 wipe that started this system, and everything rebuilt since, is documented in the build log.