XDA describes using a local Qwen model and the Pi agent on an RTX Spark laptop to build four features, including screen search and captions, in about five hours and 42 minutes. After individual feature tests passed, integrated operation and restarts exposed search delays and missing database writes. The agent fixed errors but also documented incorrect explanations of some causes. This is a personal experiment, not official Copilot+ availability or a security-validated finished product. Local AI-generated code still needs prolonged operation, restart testing and checks through the actual user interface.