0:00 What's the fastest way to stop an AI agent that's about to do something dumb? Pull the plug, right? Sure. But what happens to all the good work it was doing?

0:11 Agents don't just chat anymore. They run commands, call APIs and move money. So everyone wants a kill switch: a control outside the agent that stops it, and that the agent can't override. Good instinct. So let's look at a really good one.

0:27 On September 28th, NVIDIA announced its Open Agent Safety Platform. Part one is OpenShell. Each agent runs in a sandbox with kernel-level isolation, and the supervisor is its only way out to the network. Part two is Sentry, a watchdog on separate BlueField-4 hardware. NVIDIA says it can quarantine an agent that tries to leave its boundaries in milliseconds. Enforcement that lives outside the agent. We genuinely love that. A control the agent can argue with is just a suggestion.

0:57 Now, a thought experiment. A refund agent has handled 400 refunds today. All correct. Request 401 tries to pay an account that isn't the customer's. Hit the kill switch, and 401 never goes out. Neither does 402. Or anything else queued behind it. Work in progress stops, and someone gets to investigate and restart the agent. A kill switch asks exactly one question: should this agent keep running? For one bad refund, that's a very large hammer.

1:27 So ask a smaller question, before every single action: should this action, from this agent, for this user, happen right now? That's a control switch. Ackuity checks each action against its context: the user, the agent, its goal, the target data, the tools and the history. It compares the action with normal behavior, and asks whether a string of harmless-looking steps adds up to an attack. Then it picks a response sized to the risk. Allow. Constrain. Send to a human. Block. Or terminate. Back to our refund agent. 401 gets blocked, or sent to a person. 402 goes out on time. The agent keeps working.

2:04 And the kill switch? It stays on the ladder, as the last resort. Use it when the agent itself can't be trusted: poisoned memory, a hijacked goal, a burst of blocked actions.

2:15 Quick recap. Enforce outside the agent. Decide action by action. Allow, constrain or block the one bad step. Keep terminate for the truly catastrophic. Stop the bad action, not the whole agent. Well… not always.

2:29 Ackuity is building an integration with OpenShell's supervisor middleware, and the Ackuity sidecar runs on any Kubernetes cluster today. Want to see it on your agents? Request 30-day trial access at ackuity.ai.