The first production bill for an agentic system is often a surprise, because cost scales with autonomy. Every retry, every re-read of a document, every "let me think about that again" is billable. The fix isn't a cheaper model — it's structure.
Give every run a budget
Attach a token budget to a run the same way you'd attach a timeout. When a run approaches its ceiling, it should escalate to a human or a cheaper path rather than silently spending more. A budget turns cost from a monthly surprise into a per-case decision you can see and tune.
Spend the expensive model only where it earns it
Not every step needs your best model. Route classification, extraction, and formatting to small fast models; reserve the frontier model for genuine reasoning and ambiguity. Cache what's deterministic. The goal is a workflow where cost tracks difficulty, not habit.