'Why Software Factories Fail': Faros AI Data Shows +242.7% Production Incidents Per PR Since AI Coding Adoption

HumanLayer / Hacker News·high signal

HumanLayer founder Dex argues that lights-off AI software factories fail not because harnesses are badly configured but because models cannot maintain codebase quality over time — SWE-bench-style benchmarks reward passing tests with no penalty for eroding maintainability, and architectural decay takes months to surface, too slow to train on. He cites Faros AI data showing +25% review comments, +31.3% of PRs skipping review, +242.7% production incidents per PR, +57.9% monthly incidents, and +54% bug rates per developer since AI tool adoption accelerated. HumanLayer's own July 2025 lights-off production experiment ended within weeks after repeated incidents; 377 points, 264 comments.