The other day I heard about the concept of a software "dark factory".

How it works is no humans ever touch the code. A process engineer writes the factory "process", and a spec guy rights the software specs. A handful of AI agents will then automatically analyze the specs, write the code, review the code, test the code, and deploy/monitor the code. Bugfixes will go through the same process. A human will perform final user acceptance testing.

There are already active companies using this process. I have not used their products (that I know of). However, based on my experience with AI capability as of late 2026, I believe such an AI software factory is capable of writing high complexity, medium scale, medium reliability, consumer-grade software. Which is most SaaS out there.

Let's formalize this from a management and jobs perspective.

A traditional engineering team may compose of this:

Upstream:

• 1 product manager/business analyst
• 1 designer

Midstream (software engineers):

• 1 staff engineer
• 2 tech leads/engineering managers
• 5 regular engineers

Downstream:

• 1 devops/devex guy
• 1 QA guy

That's 12 people.

An AI software factory with roughly the same output would compose of this:

• 1 product manager/spec writer
• 1 process engineer
• 1 "foreman"/QA
• 5-10 AI agents of various types - I happen to have some insight into token usage and costs. I think with the agents running at full tilt all day, each agent will consume on the order of $100s worth of tokens per day.

That's 3 people, plus ~2 people's worth of costs going to "renting" a capital good.

So at constant output, 3 out of 4 jobs are eliminated. Token costs are probably going down in the long term. I don't think Jevons paradox will apply here, because software is already so cheap.

This is very analogous to the relationship between automation and manufacturing. The biggest conceptual difference is that AI development is happening much faster. Another difference is that a software factory produces a one instance of a product, whereas a hard factory manufactures many of a product. But this is a pedantic difference not relevant to the broader implications.

I will probably discuss the broader implications of the software factory phenomenon, and AI itself, in another entry.