FAILURE MODES / 2 AM PAGER

This AI Agent Has No Off Switch

An AI agent stuck in a retry loop isn't broken. It's doing what it was told. Nobody told it when to stop. The off switch that holds lives in the warehouse, not the prompt.

2026-11-11

An AI agent stuck in a retry loop isn’t broken. It’s doing what it was told. Nobody told it when to stop.

The 35-second reel: this AI agent has no off switch.

The prompt is not an off switch

Prompts are suggestions. An agent reads them, and then it reads them again, and then it retries. Nothing in the prompt decides when enough is enough. A turn limit in the framework helps, and you should use it. But the agent can argue with a turn limit. It cannot argue with a suspended warehouse.

The fix: put it in the warehouse

In Snowflake, the off switch that holds is a resource monitor on the agent’s own warehouse. Three steps:

  1. Give the agent its own warehouse, separate from everything else.
  2. Create a resource monitor with a daily credit quota and SUSPEND_IMMEDIATE at 100 percent.
  3. Assign the monitor to that warehouse.
CREATE RESOURCE MONITOR agent_cap
  WITH CREDIT_QUOTA = 5
  FREQUENCY = DAILY
  START_TIMESTAMP = IMMEDIATELY
  TRIGGERS ON 100 PERCENT DO SUSPEND_IMMEDIATE;

SUSPEND_IMMEDIATE suspends the warehouse and cancels running queries. Plain SUSPEND lets them finish first. Pick the one that matches how much damage a runaway agent can do while it finishes.

Fine print, from the Snowflake docs

Only ACCOUNTADMIN can create a resource monitor. Suspension isn’t instant, even with SUSPEND_IMMEDIATE, so set the quota below what you can afford to lose. Monitors cover warehouses only. Serverless features and AI services need a budget instead. The daily quota resets at 12:00 AM UTC.

The rule

Use the framework’s turn limits. Then put the real limit in the platform, where the agent can’t talk its way past it. Where does your agent’s off switch live?

This post started as an Instagram post →

Discussion

Talk it through

Argue with us on Instagram.