The AI agent paradigm is promising but current implementations are fragile. The failure modes compound — each step in a chain has some probability of error, and multi-step chains amplify this exponentially.