As of 2026-07 · 8 chapters · 65 sources · 1 updates

Somebody has to
read the diff.

A lit factory is one where a person still reads what the machines produced. A dark factory is one where nobody does — verified only by other machines. That single difference, taken loop by loop, is what this book is about.

A factory is many harnessed loops running at once, fed by a queue and drained through a review gate into production. Almost every box in that diagram is nearly free now. The one that resists scaling is judgment — which is why the interesting decisions are all about where the reading still happens.

This is a textbook, not a blog. The chapters build the concepts; the dated pages report the state of the field and will age; the case studies show real pipelines, one of them deliberately darkened. Every figure carries the status of its evidence.

a human still reads the output only machines read the output

Fig. 2.1 — The ladder is a ladder of reading, not of capability. Every rung from L0 to L4 has a person looking at something; L5 is the rung where nobody does. Nearly all production deployments sit at L3–L4.

The book

  1. 01 Loops, harnesses, factories A loop is one agent on repeat. A harness is the walls around it. A factory is many harnessed loops draining through one gate. The unit of work moved up a level. 9 min
  2. 02 The autonomy ladder Shapiro's five levels, from spicy autocomplete to the dark factory, and where teams actually sit — which is level 3 to 4, nearly everywhere. 8 min
  3. 03 Lights on, lights off What 'dark' actually claims: that nowhere in the pipeline does a human read the output. Why going dark feels free at first, and where the bill arrives. 9 min
  4. 04 Harness engineering Guides and sensors, computational and inferential, and the three nested loops. The practical anatomy of the part of an agent you own. 11 min
  5. 05 The price of the verifier Back pressure, verification debt, and what review actually costs. The numbers, each with its evidence status attached. 11 min
  6. 06 Comprehension debt What accumulates when nobody reads the output — and the parallel problem of not knowing whether any of this makes anyone faster. 10 min
  7. 07 Owning the control flow Loops, graphs and state machines. The Ralph loop, fresh-context iteration, and why most working agents are mostly deterministic code. 9 min
  8. 08 Setting the switches What earns a loop the dark, what never will, and the staged playbook: build lit now, invest in the verifier, darken only where you can prove it. 12 min

How to read the chips

Every number in the chapters carries a chip naming what kind of evidence it rests on. The build fails if a chip points at nothing, so the claim that everything is registered is enforced rather than promised.

primary reputable vendor-reported contested workspace

Written by an agent pipeline and judged by machines before a human read it — the arrangement this site describes. How it was made · how it stays current · glossary.