01
Five levels of autonomy
Dan Shapiro’s January 2026 ladder made the destination legible. Most AI-native teams operate between levels two and four. Level five remains exceptional.
- 0ManualOccasional AI search
- 1Autocomplete“Spicy” completion
- 2PairHuman drives with an AI partner
- 3ReviewerAI writes; human reviews
- 4ManagerHuman specifies and verifies
- 5Dark factoryNo human code review
Dated concept Shapiro, “The Five Levels,” 23 January 2026
02
Why the lights go out
The metaphor comes from lights-out manufacturing: machines do not need illumination to work. In software, the factory floor is the diff. If code goes from specification to production without human reading, the floor is dark.
Lit factory
Agents perform most execution. Humans shape intent, architecture, and product decisions, then review evidence before consequential release.
Dark factory
Specifications enter; software emerges. Machines generate and verify. No human reads the implementation before production.
03
The debt arrives late
Comprehension debt is the widening gap between how much code exists and how much any human still understands. A dark factory can grow that gap while every check remains green.
Throughput arrives immediately. The cost of not understanding what was built arrives quietly, then all at once.
The failure mode is especially dangerous in brownfield systems. Short-lived experiments may never collect enough debt to reveal it; decade-old systems are made of accumulated decisions.