How to Stop an AI Agent From Emailing 400 Customers at 3AM: Budget Caps, Idempotency & a Kill Switch in n8n

Описание и применение

Внедрение паттерна защиты в n8n, включающего ключ идемпотентности, контроль месячного бюджета, ограничение скорости и kill‑switch, предотвратило массовую рассылку 400 неверных писем и обеспечило контроль расходов.

Задача

An unattended AI agent sent 400 incorrect emails at 3 AM, causing operational chaos.

Результат

The guard pattern stopped unintended mass emailing, enforced budget limits, and added a kill switch, reducing operational risk.

Рабочий процесс

• 1. Generate SHA‑256 idempotency key from client, action type, recipient, payload.
• 2. Insert key into guard_actions ledger in Postgres.
• 3. Query ledger for monthly cost per client and compare to cap.
• 4. Count actions in last hour and compare to rate cap.
• 5. If any check fails, send Slack alert, stop workflow with Error node; otherwise proceed with LLM call and outbound action.

Доказательства

put one guard step in front of every LLM call and every outbound action

Источник

SHA-256 of client, action type, recipient, payload

Источник

SUM(cost_usd) by client and month vs cap

Источник

COUNT(*) of actions in last hour vs cap

Источник

Slack alert, then Stop and Error node

Источник