Software Factories Shift to Lit Models as 3-to-10-Step Loops Beat Full Automation
Updated
Updated · O'Reilly Media · Sep 18
Software Factories Shift to Lit Models as 3-to-10-Step Loops Beat Full Automation
1 articles · Updated · O'Reilly Media · Sep 18
Summary
Human judgment is moving back into AI-driven software factories, with the article arguing that fully automated “dark” pipelines create comprehension debt and eventually outrun what teams can reliably verify.
The bottleneck is verification, not code generation: agents can produce changes cheaply at scale, but review gates remain the costly constraint, so autonomy should expand only where checks are fast, reliable and hard to game.
Short loops are the best candidates for lights-out automation—Horthy’s rule of thumb is 3 to 10 steps—while longer runs past 20 steps tend to lose context and hide defects.
Lit factories keep people upstream on product, design and architecture, using familiar practices such as strong types, test seams and clear boundaries to make agent output safer before code reaches production.
The broader shift is from writing individual diffs to designing and supervising the factory itself, with engineers owning the outer loop of approval, diagnosis validation and risk-bearing.
If human verification is the ultimate bottleneck in AI software factories, will we eventually be forced to trust AI to verify other AI?
As AI writes code faster than humans can read it, how will organizations measure and survive the silent accumulation of comprehension debt?
With pull request sizes exploding in 2026, are developers permanently transitioning from creative builders to mere editors in an automated assembly line?