A dark factory is a software factory where agents produce nearly all of the code. The name borrows "lights-out" manufacturing: the line runs without people on the floor, while people still choose what gets made and whether it ships.
The closest documented approach is OpenAI's. Its harness engineering account describes an internal repository written almost entirely by agents — people prepared specifications, tools, feedback, and repository structure rather than writing code. Its Symphony spec shows the orchestration layer: a work board drives isolated agent runs with proof-of-work requirements and a protected landing path. OpenAI's internal usage figures describe adoption across engineering and other departments; those numbers are first-party reports, not an audit.
StrongDM's software-factory account is the second notable public description of an organization rebuilding its development process around agent-produced work.
What "dark" means — and does not mean
Dark describes where the code comes from, not where authority lives. In every documented case, people keep the same named gates: they choose the product work, set permissions, review what the system produces, and decide what merges and releases. "Nearly all code written by agents" is a claim about generated lines, not about unattended decisions.
It is also the end of a spectrum, not a category. A team running one bounded production line and an organization generating most of its codebase through agents differ in degree along the same variables: how much work agents touch, how much checking is automated, and how much attention people spend on direction versus implementation.
What a dark factory still needs from people
The darker the floor, the more the human work concentrates: task selection and acceptance, the specifications and context agents receive, permissions and spend limits, review of what the system reports, and the decision to expand or contract the system's scope. The documented cases invest heavily in exactly those surfaces — repository-legible context, enforceable architecture, and proof loops — because the factory's quality is decided upstream of any single run.
How to read dark-factory claims
Apply the same evidence labels everywhere. First-party accounts describe what an organization reports about itself; they are real evidence with a known bias toward the flattering number. A "1 million lines" or "no human code" figure describes generated volume, not correctness, maintainability, or business result. The questions worth asking of any dark-factory claim: what fraction of merged work is agent-produced, what checks gate it, who can stop the line, and what the system costs per accepted result.
For the orchestration design behind the best-documented case, read OpenAI Codex Symphony. For reported scale across organizations, see Agent scale.