Vibe coding observability: what to verify before production

A team merges an order export mostly written by a coding assistant; tests pass and the endpoint returns 200, yet weeks later a customer finds missing rows. The article argues the real question is what proves production behavior, not whether the AI writes good code. It rebuilds a scenario where generated code runs one query per order and swallows exceptions, producing 41 spans and silently dropping 10 orders. The author concludes that spans should sit close to outgoing calls, where generated code cannot hide errors, and that broad exception handling without logging should be banned in review. The piece also notes that its DQL queries and Dynatrace setup were not verified against a live environment.