AI Agents Cause Trouble—Who Answers the Phone?

If your AI agent causes trouble in production tonight, who gets the call? If you have to think about it before you can answer, you have a serious problem.

AI agents—AI systems that can call tools and complete tasks on their own—are already running in the operations of many companies. Franklin Solum mentioned in a tweet that Korea's SK Telecom manages every AI agent like an employee. They have an employee ID, role, supervisor, token budget, and work history.

The ID is not the point. The key is that there must be a named person accountable for the agent's failures—a specific person, not a team.

Solum suggests that every agent in production should have a card like this:

1. Owner: a real person, not a team
2. Purpose, and which systems it can touch
3. Model and prompt version
4. Tool allowlist, with write permissions flagged
5. Spend and concurrency limits
6. Date of the most recent evaluation
7. Emergency stop switch

If the owner leaves, the agent cannot keep running under the old account. Reassign an owner that same day, or turn off its write permissions.

This logic isn't complicated. The core is to give every agent a specific person who is responsible. You can scale agents without limit, but you cannot scale a system that no one is accountable for.

This image is a diagram of the agent governance process; save it and post it next to your monitoring dashboard.