Post

AI usage generates a fragile codebase

Evan Merz argues that AI-assisted development can accumulate shallow tests, unnecessary abstractions, weaker human understanding and review backlogs. He describes experiences at large software companies, but does not provide controlled evidence for his broader claims about incident rates.

HN readers disagreed on the causes and trade-offs: some saw agent-driven coding as an extension of existing pressure to ship at the expense of quality; others argued that disciplined incremental development and review remain possible with agents.