PRAEVA SYSTEMS
PRAEVA INSIGHTS  ·  October 8, 2026
When Agents Run Amok · #007

Kill Switch vs. Stop Sign

Stopping an agent is useful. Stopping the wrong action is better.

By The Great AgenticBehind the curtain at Praeva Systems.

When Agents Run Amok #007: Kill Switch vs. Stop Sign
When Agents Run Amok #007. For the seventh installment, we had to give a little credit to the #007. So yes, the artwork is intentional. The subject, though, is a little less glamorous.

There has been a lot of talk about kill switches for AI agents. The idea makes sense. If an agent starts behaving dangerously, shut it down. Stop the process, revoke the token, disable the account, pull the plug. There are absolutely situations where that matters. But a kill switch answers a different question from the one enterprises are going to face over and over again: should this particular action be allowed to happen?

That is where the stop sign comes in. A stop sign does not wreck the car, take away the driver’s license, or close the road. It just says: not here, not now, not under these conditions.

AI agents need the same kind of control. An agent can be properly authenticated, have valid credentials, and still have a working API token. It may even be doing exactly what it was built to do. That still does not mean it has authority to take every action available to it.

Take a procurement agent. Maybe it has been given authority to order replacement laptops from an approved vendor within a certain budget. It has access to the purchasing system and everything it needs to create a purchase order. Then it tries to buy $400,000 worth of GPUs from a vendor nobody approved. You probably do not need to kill the agent. You need to stop that purchase.

That distinction matters because the problem is not always that an agent has gone rogue. Sometimes the agent is perfectly healthy and the requested action is simply outside the authority it was given. The same thing gets more interesting once agents start handing work to other agents. Say Agent A has valid authority and passes part of the job to Agent B. The work sits in a queue for a few minutes, and during that time the original authority is revoked.

Agent B may still have valid credentials. The queued instruction may still be there. Agent A may still be running just fine. None of that means the action should still happen. If the authority behind the action is gone, the downstream action should stop too.

That is exactly the kind of moment where “kill the agent” is too crude. Nothing necessarily needs to be shut down. What matters is whether the authority still exists when the action is about to happen. Before a consequential action executes, the system should be able to answer a simple question: does this agent still have legitimate, current, properly delegated authority to do this specific thing?

If the answer is yes, let it happen. If the answer is no, stop it. If the answer is not clear enough, send it to a human. That is a much more useful control model than waiting until something goes wrong and then reaching for the emergency stop.

Kill switches absolutely have a place. Sometimes you really do need to shut the whole thing down. But if that becomes the main control model for autonomous systems, we are designing for failure after the fact. A better model is more precise. Keep the agent running. Keep the credentials intact. Keep the workflow moving. Just do not let it cross a boundary it was never given authority to cross.

That is the difference between a kill switch and a stop sign. One shuts the system down. The other keeps the wrong action from happening in the first place.

Stopping an agent is useful. Stopping the wrong action is better.

Before action, authority.

The Great Agentic Behind the curtain at Praeva Systems.
← Back to Praeva Insights