tl;dr: Developers are increasingly building "harness loops" around coding agents—outer loops that queue work, judge results, and re-prompt models until tasks complete without human intervention. The author argues this works well for ephemeral tasks like porting, benchmarking, and security scanning, but produces overly defensive, hard-to-understand code when used for long-lived codebases, creating systems treated more like organisms than deterministic machines. Opting out may not be viable due to competitive and security pressures, so the real question is how to preserve human judgment, code legibility, and engineering standards in a loop-driven future.