Reality Check · Ins7ghts Daily, Wednesday 4 November 2026
Two vendors sold control over AI agents on Tuesday, and they did not sell the same product.
CrowdStrike named a feature the AI Kill Switch and put it in Falcon Guardian. GrowthBook, on the same day, shipped version 5.2 with human in the loop for agents. Two different companies, in two different markets, both answering the agent question with an interruption: a person or a button gets in the way before the thing acts.
That is the claim worth testing, because it is the one every vendor is comfortable making. Control means you can stop it.
Now the counter-event, and it came from the same week and partly from the same press release. Microsoft's new Data Security Posture Management experience sells a single view of every AI app and agent active in a tenant in the last thirty days, the ones on Agent 365 included. CrowdStrike, in the announcement that carried the kill switch, agreed to buy Fiddler AI for AI agent observability, and put Falcon Guardian AI Detection and Response natively inside Anthropic's Claude Cowork rather than watching it from outside.
Read those three together and the switch stops looking like the hard part. Two vendors spent the same day on naming: what is running, where it is running, what it touched. One of them paid for a company to get it. The other needed its sensor to live inside the agent's own product, because from the perimeter the agent is a single well-behaved process.
The best case against me: containment before accounting is a defensible order of operations. When you do not know the blast radius, a button that halts the next action is worth more than a report about the last thousand. A kill switch is also cheap, fast and easy to explain to a board, and none of those are bad properties in a control.
The rest of this piece is for subscribers.
Ins7ghts Daily is free. One letter a morning from Yves Mulkers on what moved in data and AI, and why it matters to your side of the table.
Already subscribed? Sign in at the top of the page.