Production incidents per PR jump 242.7% since AI coding adoption

Production incidents per pull request climbed 242.7% since AI coding adoption took hold, per Faros AI data cited in HumanLayer's new essay, Why Software Factories Fail. PRs skipping review entirely are up 31.3% too, meaning the on-call engineer, not a reviewer, is often the first human to see the code.

HumanLayer isn't only citing outside numbers. The company ran its own lights-off, autonomous factory in production in July 2025 and killed it within weeks after repeated incidents. The essay argues the failure is structural: SWE-bench-style benchmarks reward passing tests with zero penalty for eroding maintainability. Architectural decay takes months to surface, far slower than any training signal can catch.

A second essay, from Warsaw developer Piotr, did even better on Hacker News: 763 points, 586 comments. He argues the cause is organizational, not technical. Teams point frontier models at features because stability has no KPI. He cites no data and says so. His examples are boring on purpose: a banking app that demands repeated FaceID checks before it shows a security confirmation. An LG warranty form's error is visible only in the dev console.

The Faros numbers are correlational. Adoption coincided with plenty else changing, and incidents per PR rise mechanically if PR volume outpaces incident volume. Still, the direction matches what I see in my own work. It's easier now to produce a large, plausible, compiling change that nobody, including me an hour later, understands.

The fix worth taking from this cluster comes from the essay Codegen Was the Tutorial Level. The definition of done needs to live inside the prompt itself, not a checklist you'll forget. Tests, observability hooks, operator docs, and rollback notes go in as required output. A harness that defines done as code compiles will flood you with debt that compiles fine and explains nothing.