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

Description and use

AI agents handle routine tasks well but lack business context, causing critical failures and alert fatigue. Switching to assisted mode with human approval improves reliability and speeds incident resolution.

Problem

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

Outcome

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%.

Workflow

• 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

Evidence

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

Source

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

Source

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

Source