Researchers find letting AI coding agents write their own tests backfires Startup Fortune