The Software Factory Was Built for Humans. AI Needs Something Else.

I built the software factory.

Agent pipelines. Role definitions. Detailed specs. Review gates. The whole thing. On paper it looked great. Code came out fast. PRs appeared. Tests went green.

But the longer I ran it, the more it felt off.

The entire model is still organized around human limits. We nest folders because we need spatial cues. We cling to DRY and clean-code rituals because every line used to be expensive for a person to write and maintain. We treat seniors as the only ones who can "keep the AI on the rails" — which often just means they use it as faster autocomplete for the same patterns they've written for years.

LLMs don't work that way. They care about meaning, not folder depth. They can throw away or regenerate hundreds of lines in seconds. The cost equation changed, but our mental model didn't.

The factory idea sounds modern, but it's an old industrial metaphor. Spec in, product out. Assume the design is already solved. Eliminate variation. Formalize requirements early so the line keeps moving.

Software was never manufacturing. The source code is the design. What agents actually accelerate is the design loop itself: try something, see what happens, adjust. Forcing that into an assembly line just makes us do the wrong things faster.

After living inside these systems, one thing is clear to me:

This model was designed for humans, not for AI.

We need a different frame. Something closer to design collaboration at machine speed. Less about throughput, more about clear intent, taste, verification, and knowing when to explore versus when to lock things down.

The software factory helped a lot of people get started with agents. It was a useful temporary scaffold. But it's not the right long-term model.

Time to invent the next one.