chapter 02 · the dichotomy
Light & dark
The vocabulary comes from manufacturing. FANUC has run robot factories in Japan since 2001 with the lights physically off, because only machines work the floor and machines do not need light to see; Xiaomi followed with a phone plant in 2024. “Lights-out” is not a metaphor there. It is an electricity bill.
Applied to software, the floor is the diff. A dark software factory ships code that no human has read, verified only by other machines. The word carries no judgment; it locates the reading, and in a dark factory the reading happens nowhere. A light factory (or lit factory) is the same pipeline with the lights left on where judgment lives: agents still do most of the building, but a person reads what comes out before it ships.
§ 2.1The five levels
The framing most of the industry now uses comes from Dan Shapiro’s essay “The Five Levels: from Spicy Autocomplete to the Dark Factory” (danshapiro.com, 2026-01-23, amplified by Simon Willison five days later), modeled deliberately on the NHTSA/SAE driving-automation levels:
Shapiro places himself at level 4 and notes that only a handful of small teams have reached level 5. That distribution held through the first half of 2026: nearly every production deployment with a name attached — Stripe, Spotify, Cursor — runs at level 3 or 4, autonomy in execution but not in approval. The documented level-5 cases can be counted on one hand, and chapter 5 examines the best-known of them.
§ 2.2Comprehension debt
Going dark feels effortless at first: the review step everyone resented is gone, and the throughput chart bends upward so steeply it looks mis-scaled. The costs are real but deferred, and they have a name: comprehension debt, the widening gap between how much code exists and how much any human still understands. A dark factory does not pay this debt down; it takes it on as fast as it can, with the tests green the whole way.
The cautionary tale is first-person. Dex Horthy of HumanLayer reports running a fully automated code factory for about four months with no human reading the output measured (AI Engineer World’s Fair talk, 2026). The failure that ended the experiment required painstaking manual debugging just to locate, because by then no one held a map of the system in their head. The tradeoff underneath is easy to state and easy to miss: maximizing token utilization — the number on the dashboard that looks like progress — quietly minimizes the share of the system any human still understands.
The reckoning is not dramatic. Weekend toys never live long enough to collect the debt, which is why demo videos prove little; decade-old brownfield systems are made of it. If your codebase is meant to outlive the quarter, comprehension debt is a line item whether you book it or not.