What is a dark software factory?
A dark software factory is a delivery system where AI agents carry routine engineering work from a ticket to production, and people set the rules and handle the exceptions instead of doing every step themselves.
The term comes from manufacturing. A lights-out factory runs without people on the floor: machines do the work, and engineers design the line, watch the instruments and fix what breaks. Applied to software, the machines are AI agents, the line is a pipeline of specification, implementation, verification, review and release, and the instruments are metrics, traces and quality scores.
It is different from an AI coding assistant. An assistant helps one developer inside an editor. A dark factory owns the whole path: it picks work up from the tracker, splits it, runs agents in isolated environments, checks their output, asks people for decisions it is not allowed to make, ships the change and reports what it cost.
Autonomy is a dial, not a switch
Nobody should go from zero to dark in a day. Teams usually move a project through three levels as the evidence builds up:
- 1
Assisted
Agents draft every step. A person approves at each gate: spec, design, review, merge and release. This is how you calibrate a new repository and learn where the agents are reliable.
- 2
Supervised
Agents run the line on their own and people approve only the merge and the production release. Most teams spend the longest time here.
- 3
Dark
No person on the line by default. Work stops for a human only when a change touches a protected path or matches the approval policy, or when something fails past its retry budget.
What a dark factory is made of
- Intake
- Work arrives from trackers, chat and monitoring, and a triage step decides which pipeline it belongs to.
- Pipelines
- Versioned, configurable sequences of steps with parallel branches, loops back on failure and gates.
- Agents
- Roles with a model, a prompt, tools, permissions and a budget. Coding agents such as OpenHands or Claude Code do the hands-on work.
- Sandboxes
- Isolated environments per run, so agents can build and test without touching anything they were not given.
- Verification
- Tests, CI, security scans, preview checks and an LLM judge scoring the result against a rubric.
- Gates and policy
- Rules for when a person must decide, and the controls to stop everything at once.
- Measurement
- Lead time, DORA metrics, cost per task and escaped defects, computed from the runs themselves.
When it makes sense
Dark factories pay off on work that is frequent, well specified and verifiable: dependency upgrades, bug fixes with a reproduction, API endpoints that follow an existing pattern, migrations with tests, documentation, support answers and content with a fact-check step. They pay off less on novel product design or architecture with no clear definition of done.
The prerequisite is not a better model. It is a test suite, CI, and a written definition of good, because the factory can only be as autonomous as its checks are trustworthy.
Nox Kit
Nox Kit is an open-source, self-hosted dark software factory. It provides configurable pipelines, agent runtimes, per-run sandboxes, an LLM gateway with budgets, an MCP tool gateway, approvals in Slack and Telegram, and the metrics to decide when a project is ready for the next autonomy level.