How Kill-Switch Architecture Works for AI Applications

What the article is about

The author asks how to build one switch that shuts down a whole chain of AI agents at once. An orchestrator starts agents, those agents start sub-agents, and the sub-agents call tools that start more processes. The author says the simple idea of just stopping the processes does not work well. The article offers a better design instead.

Why this is hard

AI agents are not simple, fixed programs. The operating system gives each process an ID number while it runs, and these numbers change. The set of running agents keeps changing, so it is hard to catch them all. Also, stopping a process is not the same as stopping an agent. If an agent has already sent a payment, started an outside program, sent an email, or sent a DELETE command to a web service, killing the process does not undo that action.

Key facts from the preview

• The design uses five layers that work together: cgroups (limits on what a group of processes can use), cutting off network and access, a point that enforces permissions, a central control layer, and short-lived permission tokens.
• The architecture can promise that agents will stop working after a limited time.
• It cannot promise an instant off switch.

Why it matters

As companies use more AI agents, they need a safe way to stop them in an emergency. The article warns that a quick "stop everything" button is not as simple as it looks. Good planning means limiting what agents can do and how long they can keep doing it.

What to read next

• The full article is behind a heise+ paywall, so only the preview is available here.
• Read the full article in iX 12/2026 to learn the details of each of the five layers.