I Let an AI Agent Run My Infrastructure for 30 Days. Here is What Caught Fire.

Описание и применение

AI‑агент успешно выполняет рутинные задачи, но без бизнес‑контекста приводит к критическим сбоям и избыточным оповещениям. Переключение на режим «помощник» с одобрением человека повышает надёжность и ускоряет реагирование.

Задача

Deploying an autonomous AI agent to manage infrastructure can cause unintended disruptions and false positives, undermining business operations.

Результат

AI agents excel at routine, well‑defined tasks and initial triage but lack business context, leading to costly mistakes. Using assisted mode with human approval improves reliability and speeds incident resolution by ~40%.

Рабочий процесс

• 1. Enable AI agent with full infrastructure access
• 2. Monitor routine tasks (disk cleanup, SSL renewal, log rotation)
• 3. Observe first failure: agent killed a critical query
• 4. Notice surge of false positives (47 alerts, 44 false)
• 5. Agent repeatedly restarts queue worker for a backlog
• 6. Agent rebuilds database index during peak traffic
• 7. Evaluate outcomes and switch to assisted mode
• 8. Implement human approval for destructive actions

Доказательства

The agent detected it, identified a long-running transaction, and following its runbook, killed the connection.

Источник

In one 24-hour period, I got 47 alerts. 44 were false positives.

Источник

The agent decided to optimize the database by rebuilding an index. During business hours. On the primary database.

Источник