Dark Factory AI: Agents on a Delivery Line
Short answer
Dark factory AI is the agent work on a lights-out delivery line: drafting a bounded change, running the checks you already trust, and queueing it for a person. The factory is the production system. The AI is one station on it. Dark factory AI is not a license to merge, deploy, or invent product decisions with nobody accountable.
Last updated: September 22, 2026.
Key takeaways
| Question | Answer |
|---|---|
| What is dark factory AI? | Agents that draft, check, and queue routine changes on a lights-out line. |
| What is it not? | Unsupervised merge, deploy, or product judgment. |
| How is it different from a dark factory? | The dark factory is the production stance. The AI is the station that runs the routine pass. |
| How is it different from agentic AI? | Agentic AI is the behavior. Dark factory AI is that behavior placed on a delivery line. |
| Who is accountable? | A person who can reject the change and whose name is on the merge. |
What is dark factory AI?
Dark factory AI is not a second definition of a dark factory. The production stance, what can run without a person at the keyboard, is what a dark factory is in software. The phrase "dark factory AI" names the agents on that line.
Those agents do three things worth separating.
- They take a job that already has an edge: a report, a failing check, or a written acceptance note.
- They use tools the line already has: the repo, the test command, the issue tracker.
- They stop at a queue. The merge is not their last step.
The industrial origin of the name is on what a dark factory is in software. Here the point is narrower: in software the agents are the machines, the floor is the routine pass, and the company still has owners.
IBM's description of agentic AI is the behavior underneath: a system that pursues a goal with limited supervision, using tools. Dark factory AI is that behavior given a station and a stop rule. It is not a claim that supervision is gone.What the agents are allowed to do
| Action | On a dark factory AI station | Stays with a person |
|---|---|---|
| Read the job and the repo | Yes, when the job is written down | If the job is still a debate |
| Draft a bounded diff | Yes | If the change rewrites a product decision |
| Run the checks you named | Yes | If there is no check, do not pretend there is |
| Open a review | Yes | Accepting the review |
| Deploy, delete, or pay | No | Always, unless you have a separate rule |
Anthropic's note on building effective agents draws a useful line: a workflow follows a path you defined, and an agent directs more of its own process. A dark factory station can be either. The line still ends at a person when the action is hard to undo.
How dark factory AI differs from the neighboring terms
Agentic AI can live in a chat window. Nothing about it requires a delivery line. Dark factory AI is agentic behavior, or a simpler script, assigned to routine delivery steps.A software factory is the whole production system. Dark factory AI is one way to staff the draft and check stations. A factory can also be staffed by people.
Cloud agents answer where the process runs. Dark factory AI answers whether a person is typing the routine pass. A hosted runtime can keep a job going when a laptop is closed. That does not make the merge automatic. Agent swarms are several roles on those stations, with one review gate. xSquad's delivery offer on xsquads.ai is that arrangement. This page does not attach a rate or a price to it.How to place dark factory AI on a line
Put the agent where a check can say no. If nothing can say no except another model, you do not have a station. You have a conversation.
Write the job so a stranger could reject it
The agent should be able to tell a wrong result from a right one using the note you wrote. Taste and strategy do not belong in that note. The full job spec is on the dark factory coding page.
Point it at tools you already run
The repo, the test command, and the tracker are enough to start. A new platform is optional. The stop rule is not.
End the agent at the queue
The handoff is a diff and the check result, waiting for a person. Merge, deploy, and anything you cannot undo are outside the agent's last step.
Read what it got wrong
The first useful output is the rejection. Fold that into the next job note. That is a person improving the station, not a claim that the system trains itself.
Remove it if the job is not bounded
If every run needs a new definition of done, the agent is in the wrong place. Put a person on that work, or write the job more tightly before you try again.
Frequently asked questions
What is dark factory AI?
Dark factory AI is the agent work on a lights-out delivery line. Agents draft a bounded change, run checks you already trust, and queue the result. A person still owns the merge.
Is dark factory AI the same as a dark factory?
No. A dark factory is the production stance: routine steps without a typist. Dark factory AI is the agent station inside that stance. You can discuss the stance without naming a model.
Is dark factory AI the same as agentic AI?
Agentic AI is a behavior; the factory terms are separated on what a software factory is. Dark factory AI is that behavior placed on a software delivery line, with an explicit stop before merge.
Can dark factory AI ship to production by itself?
Not on the definition used here. Shipping, deleting, and other hard-to-undo actions stay with a person unless you have written a separate rule. The agent queues work. It does not accept it.
Do the agents replace engineers?
They take routine drafting and mechanical checks. They do not replace the person who sets the job, rejects a bad diff, or owns the outcome.
Does dark factory AI need cloud agents?
No. Cloud is where the runtime sits. Dark factory AI is what the agents are allowed to do. A hosted runtime is a fit when the job should continue after a laptop closes. It is not a requirement for the idea.
What should the agent not decide?
Anything you cannot write a check for. The production bar for what must stay with a person is on what production means for dark factory software.
How is this different from a coding assistant?
A coding assistant helps a person who is at the keyboard. Dark factory AI runs the routine pass and stops at a queue, so the person starts from a diff instead of from a blank file.
When is the phrase just marketing?
When the page promises an empty building, a speed, or a headcount change. The durable meaning is a station with a stop rule. Anything past that is a separate claim this page does not make.
Which page should I read for the definition of dark factory?
Read the dark factory page for what stays human, and the software factory page for how the term sits next to code factory. Stay here if you are deciding what the agents on that line are for.
Sources
- Lights-out manufacturing for the industrial metaphor.
- IBM on agentic AI for the limited-supervision behavior underneath the station.
- Anthropic, Building effective agents for workflows versus agents that direct their own process.
- What is a dark factory in software? for the production stance.
- What is a software factory? and what is agentic AI? for the neighboring definitions.
- xsquads.ai and how we run agent swarms for a staffed review gate.
No throughput, uptime, or savings figures are claimed here.
If you want agents on that line with a person on review, 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.