Token costs are decided by architecture, not prompt wording. A production playbook for model routing, prompt compression, context pruning, and caching — so your AI employees stop leaking money on every turn.