Why does my agent keep spending after any sane budget would have cut it off?

AI agents running multi-step workflows consume tokens at 20 to 100 times the rate of a single query, and there is no infrastructure-level primitive to enforce a spending or compute ceiling at runtime. When two agents enter a recursive clarification loop or a retrieval agent over-fetches context, the only signal is the invoice at the end of the month. Frameworks like LangGraph and AutoGen each handle retries and checkpoints but none enforce a resource contract that halts execution when a declared budget is breached. A January 2026 arXiv paper formalizes what such a contract would look like, but the gap between that formalism and a deployable primitive that works across model providers and tool calls remains wide open.