Engineer Formalizes 8-Stage Software Factory for AI Coding, Keeping Humans in 2 Key Steps
Updated
Updated · O'Reilly Media · Sep 4
Engineer Formalizes 8-Stage Software Factory for AI Coding, Keeping Humans in 2 Key Steps
3 articles · Updated · O'Reilly Media · Sep 4
Summary
An engineer behind Decoding AI’s Squid system mapped an AI-driven “software factory” onto an eight-stage software development workflow, arguing it is a more useful frame than newer labels like loop or graph engineering.
The model starts with triage, brainstorming and planning, moves through implementation, review, CI and release, then feeds production incidents back into intake, all supported by a shared context layer such as an LLM wiki.
Two stages still require humans most: brainstorming and planning. The article argues strong plans let cheaper models execute effectively, while weak plans can make lower-tier models costlier through repeated retries.
Squid’s first version was overbuilt and hard to interrupt, leading the author to favor smaller, controllable commands over fully autonomous pipelines unless teams can justify the added operational overhead.
The broader takeaway is to automate bottlenecks gradually, buy platforms when observability and governance become full-time work, and expect humans to remain central at least in planning even as AI writes more code.