A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments