Some uncomfortable truths about AI coding agents

A 20-year software engineering veteran makes a detailed case against using LLM-based AI coding agents in production, covering four structural concerns: skill atrophy as engineers become code-review-only supervisors, unsustainably low AI pricing disconnected from true infrastructure costs, prompt injection as a foundational unfixable vulnerability, and copyright/licensing risks. The author's position is not anti-LLM broadly — they recommend models for non-agentic tasks like documentation lookup and code review assistance — but argues specifically that agentic, autonomous code generation introduces compounding risks that outweigh the productivity gains.