tl;dr: The author compares AI coding agents to "rockstar developers"—prolific, clever, but leaving behind incomprehensible code nobody else can maintain—except worse, because LLM-generated code comes from hundreds of disconnected contexts with no unifying design. This creates exponentially growing complexity and technical debt that may be unpayable, forcing teams into dependency on LLMs just to understand their own systems. The recommendation: guide LLMs to produce small, understandable snippets, resist over-engineering, and don't be afraid to slow down or write code yourself.