video1.0<iframe src="https://www.loom.com/embed/bcef1180635c4420a8feb78a64279b87" frameborder="0" width="1920" height="1440" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>14401920Loomhttps://www.loom.com14401920https://cdn.loom.com/sessions/thumbnails/bcef1180635c4420a8feb78a64279b87-07b5d71685a23910.gif122.424This Loom explains a project for catching and correcting AI agent errors before they handle real cases in a harness testing environment. The speaker notes that failures are otherwise hard to diagnose because there is no visibility into the cause. They simulate customer returns objective scenarios where AI agents can make inconsistent decisions by skipping evidence, misleading policy exceptions, or repeating mistakes. The UI runs iterations on three scenarios, including eligible returns, delivery group misses, and defect exception, and updates two files with trace fields showing why Atlas ejects due to reaching a limit.