What Production Means for Dark Factory Software
Short answer
Dark factory software, in production, is a delivery system you can run again: a job in, a check, a human gate, and a record of what shipped. The old promise attached to this URL, building without headcount, is not a claim this page makes. Agents can take routine steps. They do not remove the need for an owner.
Last updated: September 22, 2026.
Key takeaways
| Question | Answer |
|---|---|
| What is dark factory software in production? | A repeatable line: job, check, human gate, and a record of what shipped. |
| Does this page claim you need no engineers? | No. The URL is old. The claim is not. Someone still owns the merge. |
| What makes it production rather than a demo? | A check you trust, a person who can reject, and a record after the fact. |
| How is this different from the dark factory definition? | The definition says what can run unattended. This page says what must exist before you rely on it. |
| Where do agents sit? | On the routine stations. Not on the gate, and not on irreversible actions. |
What "production" has to mean here
A demo shows an agent writing a diff. Production is the system you are willing to run on the next real change, and the one after that. Dark factory software earns that name in production only when four things exist.
- A job written so a wrong result is visible.
- A check that does not depend on the agent's own opinion.
- A person who can send the change back.
- A record of what ran, what passed, and who merged.
Miss any one of those and you have a prototype that happens to use agents. Calling it a factory does not add the missing piece.
The vocabulary sits on what is a software factory. Which steps may run without a typist sits on what a dark factory is. The stations, named as a code line, sit on what a code factory is. This page is the production bar.
The minimum line
| Piece | If it is missing | What "present" looks like |
|---|---|---|
| Job | The agent invents the task | A report or acceptance note a stranger could reject |
| Check | The only signal is the draft itself | A test, a typecheck, or a rule you already run |
| Human gate | The queue merges itself | A named person can block the change |
| Record | Next week nobody knows what ran | The diff, the check result, and the merge are findable |
| Stop rule | The agent keeps going into deploy | Hard-to-undo actions are outside the routine pass |
Why "without headcount" is the wrong test
This URL used to argue that a dark factory replaces hiring. That argument mixed up a production system with a staffing plan. They are different questions.
A line can make repeated work easier to see. It can also sit idle because review is the scarce station, or because the jobs are not bounded. Neither outcome is a headcount figure, and this page will not invent one.
Use headcount language only if you are doing your own planning, with your own numbers, somewhere else. Do not read it out of a metaphor about lights.
xSquad is a managed delivery squad: agents on production steps, a person on review. Terms and the offer are on xsquads.ai. The arrangement of agents is how we run agent swarms.
How this sits next to agents and cloud
Agentic AI is the behavior that can staff the draft station. It is not the production bar. A chat session that writes a function has the behavior and none of the four pieces above. Cloud agents are a place to run the station. Production is about the gate and the record, which you still need if the runtime is a laptop. Dark factory AI names the agents on the unattended stations. This page names what else must be true before you treat their output as the way you ship.How to tell if dark factory software is production-ready
Walk one real change through the four pieces. If you have to improvise a piece for the walkthrough, it is not ready.
Pick a change that already shipped the old way
Use a kind of fix your team has done before. A greenfield idea hides missing checks, because nobody knows what failure looks like yet.
Show the check without opening the diff
Someone who has not read the code should be able to say pass or fail from the job and the command. If they need the agent's explanation, the check is not mechanical.
Name the person who can send it back
If the only path is forward, you do not have a gate. Write the name before you talk about running unattended.
Find last week's record
Open the place where the diff, the check, and the merge would be. If that place does not exist, the line is not production. It is a session on someone's machine.
Keep irreversible steps off the routine pass
Deploy, delete, and payments stay with an explicit human rule. A production dark factory is stricter about those steps, not looser, because you intend to run it again.
Frequently asked questions
What is dark factory software?
Dark factory software is a delivery system whose routine steps can run without a person typing each one. In production it also has a check, a human gate, and a record of what shipped.
Does dark factory software mean no engineers?
No. Agents can draft and run mechanical checks. An owner still sets the job, rejects bad work, and is accountable for the merge. This page does not claim you can drop headcount.
Why does the URL say "without headcount"?
The address is from an older version of the article that made a staffing claim. The claim is withdrawn. The page now describes the production bar, not a hiring plan.
How is production different from a demo?
A demo produces a diff once. Production is a path you can run on the next real change: a written job, a check you trust, a person who can say no, and a record.
Is this the same page as the dark factory definition?
No. The definition page says which steps can be lights-out. This page says what must exist before you rely on that line for real changes.
What should never be on the unattended pass?
Deploys, deletions, payments, and product decisions you cannot write a check for. Those need a person even when the draft station is unattended.
Do cloud agents make software production-ready?
No. A hosted runtime is a place to run work. Production readiness is the gate and the record. You can fail both on a cloud agent, and you can meet both without one.
How is this different from agentic AI?
Agentic AI is a behavior; the factory terms are separated on what a software factory is. Dark factory software in production describes the delivery system around that behavior, including the stop. One does not imply the other.
What is a record, concretely?
The diff, the result of the check you named, and the person who merged. If a later reader cannot find those, the line did not happen in a way you can audit.
Where should I start if I am new to the terms?
Start with the software factory page for vocabulary, then the dark factory page for what stays human. Use this page when you are asking whether a lights-out line is actually fit to run.
Sources
- Lights-out manufacturing for why an unattended line still has process control.
- What is a software factory? for the vocabulary.
- What is a dark factory in software? for the lights-out stance.
- What is a code factory? for the stations.
- xsquads.ai and how we run agent swarms for a staffed review gate.
No headcount, savings, or throughput figures are claimed here.
If you want a production line with a person on the gate, start at xsquads.ai.
running a shopify store? → /shopify
Last updated: September 22, 2026. First published: April 28, 2026.
Ready to Scale Your Development Team?
Try a squad on your repo. You review every PR.