Dark Factory AI Coding: Bounded Changes
Short answer
Dark factory AI coding is code generation on a lights-out station: the agent proposes a bounded diff, a mechanical check accepts or rejects it, and a person decides the merge. Autonomous here means nobody typed the diff. It does not mean the change is unsupervised. Scale means more of those bounded jobs can sit in parallel.
Last updated: September 22, 2026.
Key takeaways
| Question | Answer |
|---|---|
| What is dark factory AI coding? | An agent proposes a bounded diff, a check accepts or rejects it, a person merges. |
| What does autonomous mean? | Nobody typed that diff. It does not mean nobody is accountable. |
| What does scale mean here? | More bounded jobs can wait in parallel. Not a speed or volume promise. |
| What is a bad first job? | A vague feature, a product bet, or a change with no mechanical check. |
| How is this different from dark factory AI? | Dark factory AI is the station. This page is the coding job on that station. |
What autonomous code generation means
Autonomous code generation, in this phrase, is a narrow event. An agent reads a job, edits the repo, and offers a diff. A check that you chose, not the agent's summary, says whether the diff is worth a person's time. A person merges or sends it back.
That is the whole event. It does not include deciding what the product should be, and it does not include deploying because the diff exists.
The wider station is dark factory AI. The production stance is what a dark factory is. The system around both is a software factory. This page stays on the coding job: what "generated" and "at scale" are allowed to mean.
IBM's agentic AI overview describes systems that act toward a goal with tools and limited supervision. Limited is the word that matters for coding. The limit is the job boundary and the check.What a bounded coding job includes
| Piece | Belongs in the job | Does not belong in the job |
|---|---|---|
| Intent | A bug, a failing test, or a written acceptance note | "Make the product better" |
| Boundary | Which paths, files, or behaviors may change | A license to wander the repo |
| Check | A command you already run in CI or locally | The agent's claim that it looks right |
| Handoff | A diff and the check result | A merge, a deploy, or a release note as fact |
| Rejection | A way to say the diff is wrong | Silent acceptance |
Anthropic's building effective agents distinguishes a predefined path from an agent that directs its own tool use. Either can generate code. Neither is finished when the file is written. Finished, here, is a check plus a person.
What scale means without a number
Scale, on a dark factory coding line, means the jobs are similar enough to share a check and a gate, so more than one can be in progress without inventing a new process each time. Parallel work is a property of the queue. It is not a measurement of how many changes you will get, and this page does not offer one.
Where those jobs run is separate. Cloud agents are a hosted runtime, which can hold a job open when a laptop is closed. The coding definition does not require that runtime. A local agent that proposes a bounded diff is still dark factory AI coding. It is just a smaller room.
More generated code is not the goal. A smaller, checkable diff is a better job than a large one that only the author can judge. If the check cannot see the change, generating more of it does not make a factory.
xSquad uses agents for that drafting step inside a managed squad, with human review before ship. The offer is on xsquads.ai, and the multi-agent arrangement is how we run agent swarms.
How to bound a dark factory coding job
Start from a change you can fail on purpose. If you cannot describe failure, you cannot generate the change responsibly.
State the failure in the job
Name the broken behavior or the failing check. The agent should not have to infer the goal from a mood.
Fence the files or the behavior
Say what may change. An unfenced "fix the app" job is how a coding agent wanders into unrelated edits.
Name the command that can fail the diff
Put the command in the job. If the diff cannot fail that command, it is not ready for a person, and it is not ready to call done.
Hand off the diff, not a verdict
The agent reports the check result. It does not declare the change accepted. Acceptance is the reviewer's step.
Split the job if the diff explains two decisions
Two decisions in one diff means the boundary was wrong. Send back a smaller job. Scale is many small jobs, not one large guess.
Frequently asked questions
What is dark factory AI coding?
It is code generation on a lights-out station. An agent proposes a bounded diff, a mechanical check accepts or rejects it, and a person decides whether to merge.
What does autonomous code generation mean?
It means a person did not type the diff. It does not mean the change skipped review, and it does not mean the agent chose the product direction.
How is this different from dark factory AI?
Dark factory AI is the agent station on a delivery line: draft, check, and queue. This page is specifically the coding job on that station, and what a bounded diff has to include.
What does "at scale" mean if there is no number?
It means similar jobs share one check and one gate, so more than one can be in progress without a new process. It is not a claim about how many changes you will ship.
Is more generated code better?
No. A small diff that a check can judge is the point. A large diff that only its author understands is a worse job, even if an agent wrote it quickly.
Can the agent merge its own code?
Not in this definition. The agent stops at a diff and a check result. A person merges. Hard-to-undo actions stay off the coding station.
Do you need cloud agents for this?
No. Cloud agents are where a runtime can live. Dark factory AI coding is the shape of the job. You can do that job on a local agent.
How is this different from a coding assistant?
A coding assistant works while a person is driving. Dark factory AI coding proposes the diff from a written job and waits. The person starts at review, not at a blank editor.
What is a bad job to automate?
If you cannot write a mechanical boundary and a check for it, it is not a coding job. The production bar for what must stay with a person is on what production means for dark factory software.
Which page defines the factory around this?
The software factory page defines the production system. The dark factory page defines what stays human. Stay here if you are writing the coding job itself.
Sources
- IBM on agentic AI for limited supervision as the underlying behavior.
- Anthropic, Building effective agents for a predefined path versus an agent directing its own tools.
- Dark factory AI for the station this coding job sits on.
- What is a software factory? for the production system.
- What are cloud agents? for where a runtime can sit.
- xsquads.ai and how we run agent swarms for a staffed review gate.
No volume, speed, or savings figures are claimed here.
If you want bounded coding jobs with a person on the merge, 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.