Time for the dark factory?

A dark factory is a real thing. Not a metaphor we invented for LinkedIn.

It is a plant where people are not allowed on the floor while production runs. Robots move at speeds that would injure a human if one walked in. The door might as well say: no humans beyond this point.

People still picture pitch black. Sometimes the lights are off. Often they stay on. "Dark" means dark of people, not always dark of bulbs. Think sealed fishbowl. Great for sterile work. Great for shipping and manufacturing where the machines go faster than a human body can safely share the space.

These already exist.

FANUC in Japan has run lights-out production for years. Robots building more robots, with almost nobody on the floor for long stretches. Foxconn has documented lights-off component lines. Zeekr's Ningbo body and welding shops run as dark zones. Xiaomi's Beijing EV plant is packed with robots (a car every few minutes) even if people still work final assembly. Lear in Michigan opened a centre to show 100% automated thermal systems. Not every site is empty end to end. The sealed cells are real. This is not science fiction. It is on the map.

Normal factories are full of signs and fences because people have to move through them. Dark factories remove that constraint. You control what goes in. You trust the automated checks inside. You care about what comes out: does it work, can it scale, is it safe and secure.

That is the idea.

We usually say software innovates first, then the physical world copies. Here the order flipped. Physical dark factories have been real for a long time. Agentic software dark factories are the late ones.

Valantis Vais pushed me into this on a thread about code comments and human gates. Fair. Once you see the automotive frame, the next step is obvious for agents too.

Input: specs, context, data quality, the tests the coding agent never sees.
Harness: gates, stop conditions, blast-radius rules, provenance.
Output: a finished, checked product. Humans validate at the dock, not by reading every line on the shop floor.

With models like GPT-6 Astra and Claude Fable 5.1 this strong, and with harnesses getting more serious, are we just delaying the inevitable?

I do not mean "turn the lights off and hope." Unreadable code is not the same as safe code. If we stop reading every line, we need stronger checks, not weaker ones. Named owners. Real holdouts. Automated gates that tell the truth.

But pretending we will forever review every agent diff by hand, while FANUC and Zeekr already run plants without humans on the floor and build machines we trust with human lives, feels like nostalgia dressed up as caution.

Physical dark factories are here.
Are agentic ones overdue?

Discuss.