Token Budgets: An Empirical Catalog of 63 LLM-Agent Budget-Overrun Incidents

arXiv·medium signal

This arXiv paper catalogs 63 real LLM-agent budget-overrun incidents and presents a case study in Rust mitigation, putting hard data behind the 'runaway agent burns tokens' failure mode. With Copilot, Codex, and Claude Agent SDK all moving to metered/credit billing this month, cost-overrun characterization is suddenly a first-class operational concern. A practical reference for engineers adding spend guardrails to agent loops.