Dark software factory: only for MVPs and PoCs
Letting agents build software with minimal human touch can be a fast way to a prototype. open-agentix offers it as an opt-in mode with a fixed notice, and keeps approval gates for production.
This post is also available in Deutsch.
A “dark factory” is a plant that runs without people on the floor, so the lights can stay off. The software version is an agent pipeline that takes a specification, writes the code, writes the tests and opens the pull request with little or no human involvement.
open-agentix supports this as an opt-in agent mode with a fixed notice:
Recommended for MVP and proof-of-concept development only. Not for production changes without review.
This short post explains why we wrote that sentence into the product.
Where it makes sense
A prototype has a short life and a small blast radius. If the goal is to find out in an afternoon whether an idea works, the cost of a missed edge case is low, and the cost of waiting for review of every line is high. For a proof of concept, a throwaway internal tool or a first version of something that will be rewritten, letting agents run from specification to pull request is a reasonable trade.
Where it does not
Production software carries obligations that a pipeline without review cannot meet by itself:
- Nobody has read it. Tests written by the same agent that wrote the code check what the agent understood, not what you meant.
- Security and maintenance are invisible costs. A dependency that should not be there, a permissive default or an unreadable module all pass the build.
- Accountability needs a person. When a change reaches customers, someone has to own it.
For these cases the platform keeps the approval gates. Merge and deploy stay behind a human decision or an explicit policy that a person wrote and signed off.
What is in the platform today
In version 0.1.0 the mode exists as an opt-in flag with the fixed notice. Around it sit the pieces that make an agent-built change reviewable instead of trusted:
- Development guidelines as versioned sets (coding standards, branch and commit rules, coverage expectations, security rules, forbidden dependencies). They resolve from global to tenant to agent, and the stricter rule wins.
- A hardening review that checks a development agent’s output against those guidelines. It is deterministic first, and it can only make a verdict stricter, never weaken a policy. Findings go to the audit trail.
- The same policy gate, budgets and audit trail that apply to every other agent.
A ready-made pipeline template for the whole path from specification to pull request, with the approvals described above, is planned for the next release. It does not exist yet.
An honest comparison
This project is itself written largely by an AI agent, so it is fair to ask whether it is a dark factory. It is not. Every change goes through a pull request that humans review, the project lead owns the decisions, and tests, coverage gates and CI run on each change. The agent writes; people decide what is merged. A dark factory removes that last step, and that is exactly the step we recommend keeping for anything that matters.
If you try the mode, treat the output as a prototype. When the prototype proves the idea, put a person and a review in front of the version you intend to run.