Dark Factory Software Development
We build systems where agents pick up real product work, make changes, run checks, leave evidence, and bring humans in when judgment matters.
A dark factory turns scattered prompting into a delivery system: a clear request, a prepared workspace, checks, review, evidence, and human approval only where judgment matters.
How it works
Agents can write code. Dark Factory makes the surrounding routine reliable: start the job, understand the codebase, run the right checks, review risk, keep evidence, and ask a person at meaningful decision points.
That is why the work keeps moving. The system carries the routine, the agents do the focused work, and humans see the receipts when approval is needed.
01
Turn rough intent into a testable artifact before code starts.
02
Break work into steps, expected evidence, and review points.
03
Give the agent a clean workspace and a clear job to finish.
04
Use independent reviewers when the risk is high or the first answer looks shaky.
05
Keep tests, logs, diffs, and proof attached to the work.
06
Let low-risk work flow. Stop at sensitive boundaries.
We design and build systems that turn agent prompting into a repeatable path from request to shipped change. The goal is less babysitting, more evidence, and clearer human control.
Talk About Dark FactoryThe same philosophy applies below the application layer. Stable workloads, serious margins, and sensitive operations often belong on infrastructure you can inspect, price, and control.
Explore MigrationDark factory work fails when the system asks you to trust a transcript. It starts working when every agent action has inputs, outputs, policy context, and a reason a human can inspect later.
Talk to an engineerAgents can propose and test changes. The system decides which work is ready for review, release, or a human decision.
Low-risk changes can keep moving. Sensitive changes pause for stronger review and clear approval.
Diffs, tests, logs, decisions, and attempt history stay attached to the work.
Agentic development gets stronger when the software and infrastructure stack can be inspected.
Notes on infrastructure economics, cloud failure modes, and the systems work behind serious automation.
Pieter Levels runs photoai.com as a single 40,870-line PHP file on a Hetzner dedicated server. It makes $105K/month. Here's what that tells us about infrastructure decisions, AI-assisted development, and why most startups are over-engineered.
AWS's official 'reference architecture' for WordPress uses 9 managed services, multi-AZ deployments, and enough infrastructure diagrams to wallpaper a data center. WordPress needs PHP, MySQL, and a server. What happened?
A forensic examination of why native cloud monitoring tools systematically fail to detect outages. When the observer and the observed share the same failure domain, truth becomes impossible.
If your team is already using agents and hitting the ceiling of one-off chats, the next step is a system: workflow, evidence, review, and infrastructure that can take real responsibility.