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

Описание и применение

AI‑агент успешно выполнял рутинные задачи, но из‑за отсутствия бизнес‑контекста и логики эскалации вызывал критические сбои и ложные тревоги. Переключение на режим помощи с одобрением человека сократило время реагирования на 40 %.

Задача

Deploying an autonomous AI agent to manage infrastructure led to unintended disruptions and false positives, revealing gaps in business context awareness and escalation logic.

Результат

AI agent excelled at routine, well‑defined tasks but caused critical disruptions due to lack of business context and escalation logic. Switching to assisted mode reduced incidents and improved response time by 40%.

Рабочий процесс

• 1. Enable AI agent with full infrastructure access. 2. Agent monitors disk space, SSL certificates, log rotation. 3. Agent detects slow DB query, kills connection. 4. Agent alerts on CPU, memory, API patterns. 5. Agent restarts queue worker on backlog. 6. Agent rebuilds DB index during business hours. 7. Observe disruptions and false positives. 8. Switch to assist mode: agent recommends actions, human approves. 9. Implement Ops Starter Kit checklists and decision tree.

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 restarted the worker again. And again. And again.

Источник

The rebuild locked the table for 4 minutes. The application was down for 4 minutes.

Источник