Addy Osmani explores the concept of 'software factories' — systems where AI agents handle code generation at scale — and distinguishes between 'light' factories (humans review before shipping) and 'dark' factories (fully automated, no human reads the code). The core argument is that the bottleneck in agentic development is verification, not generation. Dark factories accumulate 'comprehension debt' silently while tests stay green, eventually leading to systems no human understands. The solution is keeping humans in the 'outer loop' — owning architecture, design decisions, and review gates — while delegating well-bounded, verifiable inner loops to agents. Good software architecture (types, test seams, clear boundaries) becomes a safety net against agent mistakes, and the key skill is knowing which loops can safely run unattended versus which require human judgment.
