I went through 14 reports of AI agents burning money

Описание и применение

Агенты ИИ часто превышают бюджет из‑за гонки между проверкой бюджета и фактическим вызовом API. Решение – резервировать худший случай до запроса и отклонять вызовы, которые превысят лимит.

Задача

AI agents often exceed budget due to race conditions between budget checks and actual API calls, leading to excessive spending.

Результат

By pre‑reserving a worst‑case cost and refusing calls that would exceed the budget, the library eliminates the race condition that caused 14 documented failures, ensuring agents stay within budget.

Рабочий процесс

• 1. Agent reads current budget for the requested model. 2. Library reserves a worst‑case amount for the anticipated call. 3. If reservation would exceed the ceiling, the call is refused. 4. If reservation succeeds, the call proceeds. 5. After the call, the library settles the actual cost and updates the budget.
• 6. Repeat for each subsequent request.

Доказательства

Nothing re‑checks budget on the deployment that actually bills.

Источник

multiple agents can observe remaining budget and start model calls before the governor's next poll.