Software Factory Archive

Her name was Priya Anand and she was pursuing a master's degree in software engineering at Carnegie Mellon. Her thesis advisor had told her to find something interesting, and she had found something interesting. She sent an email to all three of them on the same day, with separate questions tailored to each person's public work. That alone told Justin she had done her homework.

Justin's interview lasted ninety minutes. He spoke in complete paragraphs. Priya recorded the whole thing and later told her advisor that it was like interviewing someone who had already written the textbook and was now patiently explaining it to someone who hadn't read it yet, except he never made her feel like she hadn't read it.

Jay's interview lasted two hours and fifteen minutes because he kept going on tangents. Productive tangents. He talked about distributed systems, about the philosophy of testing, about how the factory's satisfaction metrics were essentially a probabilistic reformulation of something the SRE community had been circling around for years. He used the word "hygroscopic" exactly once, in a metaphor about how good architectures absorb change. Priya wrote it down and later used it as a section header.

Navan's interview lasted one hour. He brought his physical notebook—number eight at that point—and flipped through it to find specific dates, specific decisions, specific moments when something had clicked. He described the Cedar policy system with the enthusiasm of someone describing a favorite novel. He drew diagrams on a napkin and then photographed the napkin and texted the photo to Priya during the interview itself so she wouldn't have to reconstruct them from her notes.

The thesis was 200 pages. It covered the methodology, the tooling, the architecture, the organizational structure, the philosophical underpinnings, and the empirical results. There were forty-seven figures. There was a twelve-page appendix of raw satisfaction data that Priya had been permitted to collect over a three-week observation period. There was a comparative analysis with four other AI-assisted development approaches, none of which achieved the same compounding-correctness property.

The abstract was one sentence.

"This thesis demonstrates that non-interactive, specification-driven software development using autonomous agents achieves compounding correctness when the development process is modeled as a convergent dynamical system rather than a linear pipeline."

Her advisor read the abstract, then the thesis, then the abstract again. "That's a bold claim for one sentence," he said.

"I spent 200 pages supporting it," Priya said.

"I know. That's why it works."

She sent the final draft to all three team members. Justin replied with two specific corrections and a note that her analysis of the Attractor pipeline was the clearest external description he had seen. Jay replied with congratulations and a link to a vocabulary blog he thought she'd enjoy. Navan replied with a photo of his notebook where he'd written "Someone wrote 200 pages about us" with three underlines and an exclamation point.

The thesis was accepted with distinction. Priya got a job at a startup in Austin. She built them a software factory in three weeks.