Note / 2026-10-12

A software factory that ran with the lights off left code nobody could explain in four months, and the fix was reading again

One company let agents ship small and medium work unread from July 2025. The first hard bug cost two weeks, and by November it rewrote from scratch. Its correction: a short written design, then slices a person reads.

For CTOs running coding agents on a shared codebaseViaro Networks1 min read
A software factory that ran with the lights off left code nobody could explain in four months, and the fix was reading again

The first software factory we ran with agents left code nobody could explain after a quarter. One company has now written up the same failure.

Dex Horthy, cofounder and CEO of HumanLayer, which sells a tool for coding with agents, published a two part essay on July 24 and 25, 2026; it is one company's account. His team went lights off in July 2025, with agents shipping the small and medium work unread. The first hard bug cost two weeks inside code nobody had read in three months; by the third time, in November, they rewrote from scratch. His reason: models are rewarded for tests that pass in seconds and never penalized for eroding maintainability. His correction is the one we kept: a short written design first, then slices of 100 to 200 lines a person reads before the next starts. By his count 40% of tasks still go straight to the agent, and a one shot change needs close to 50% rework.

Guillermo Rauch, CEO of Vercel, on X: reject non-understanding; the risk is that reading itself gets discounted, worn down by unverified AI output.

What stayed on our line is the reading; the plan exists so the reading is short.

30 minutes of planning saves hours of review. He wrote it; we had paid for it.

Sources