Multi-agent software factory, reviewed by the book

In a dark factory nobody reads the code, so the factory has to catch its own bad work. Ours is a team of five Claude Code agents in Band Desktop. The human sends one task per stage and nothing else. One agent builds, and three check it from different sides: a test-designer writes black-box acceptance tests from the spec without reading the code, a reviewer reviews through fresh subagents and ends each stage with a final review that runs the organisers' checks in isolated mode and probes the running build, and a product-designer owns the screens. A coordinator hands off the full spec, keeps checking rounds going only while failures fall, and settles disputes by quoting the spec. No agent approves its own work, and hooks enforce it: each agent can write only its own files, history cannot be rewritten, and no one can ask the human. The rules come from books human teams use, such as Domain-Driven Design, A Philosophy of Software Design, Designing Data-Intensive Applications and Refactoring, written into skills the agents load. Defects that slip past the tests become cases the next stage tries first. In the submitted run the agents built all four stages in 10 hours with no human message after each task. The final review sent every stage back, and 31 blocking issues were fixed before acceptance, against 1 for the organisers' minimal three-agent setup on the same model. FACTORY.md and seat-config/ hold everything needed to stand the factory up.