A contrarian view making waves on Hacker News argues that AI agents in software development represent one of the field's most expensive mistakes. The author contends that agents don't truly program—they're statistical models mimicking code distribution, producing output that's broken but increasingly difficult to detect.

Drawing from six months of hands-on experimentation, the writer tested agents on projects including tinygrad development and USB-to-PCIe chip reversal. While agents excel at rapid prototyping and serve as enhanced search tools, they consistently fail to deliver production-quality code. The author notes that Apple is pushing AI adoption across all engineering teams, raising questions about whether platforms like macOS will improve or deteriorate.

The core issue lies in organizational dynamics. High-performing developers possess error-correction instincts to identify AI-generated 'slop,' but lower performers lack this filter. Large organizations with slower feedback loops face particular risk, as bottom-tier contributors can produce tenfold more output while maintaining lower quality standards.

The piece aligns with LeCun and Marcus's skepticism about current LLM capabilities, arguing that real programming agents require world models rather than statistical pattern matching. The author concludes that this era's defining story will be which organizations avoid self-harm through AI-driven code quality collapse.