Founders usually budget for token costs and stop there. In every system we've shipped, token spend is the smallest recurring cost — usually under $200/month even at meaningful volume, because most agent tasks don't need the largest model for every step.
The real costs are retries, monitoring, and the human review layer for anything high-stakes. An agent that fails silently on a malformed API response is worse than no agent at all, so production systems need logging, alerting, and a fallback path — that engineering effort dwarfs the inference bill.
Model routing matters more than model choice. Sending every request to the most expensive model is the single most common way we see agent costs balloon. We route classification and extraction to small, cheap models, and reserve the frontier models for the steps that actually need deep reasoning.
Budget rule of thumb we give clients: expect infrastructure and monitoring to cost 3-5x the raw token spend in month one, dropping close to token-cost-only once the system's edge cases are known and handled.