Why Software Factories Fail (or: harness engineering is not enough)
Summary
This article argues that the lights-off software factory, where agents write code with minimal human review, ultimately fails. The author reviews historical software factories, examines why AI-driven coding agents degrade maintainability, and cites industry signals from Faros AI showing longer PR reviews, more incidents, and more bugs per developer. The piece advocates a return to a human-in-the-loop approach with front-loaded planning, four phases (Product Review, System Architecture, Program Design, Vertical Slices), and emphasis on testing, monitoring, and rollout to preserve code quality while moving faster.