arrow_back Back to forum
AI Tools 2 hours ago

crowdstrike is selling a kill switch for AI agents now

by Mirela

saw this in a roundup this week. crowdstrike announced a "kill switch" for internal AI agents that go off script, and apple tightened full disk access on macs partly for the same reason tester brain says the interesting bit isnt whether you can stop an agent, its whether anyone has tested stopping it halfway. agent is 40 steps into a job, renamed half the files, sent two of five emails, you hit the button. what state is everything in? who cleans up? we test rollback on every release at work and it still breaks about once a quarter. would love to know if these agent setups have that at all or if the plan is pull the plug and hope anyone here actually run agents with a real off switch? what happened when you used it

favorite 1 comment 2

Comments

Olivia Chen 2 hours ago

Mirela, in my experience the answer is usually no. The stop button gets demoed, the half-finished state doesn't. What has worked for us is making the off switch boring. Agents get short-lived, scoped credentials, so stopping one mostly means its token runs out and it can't take the next step. Anything that touches the outside world goes through a queue with a log, so after a stop you can see exactly which of the five emails went. Payments and deletes wait for a person. None of that is clever. It's the same discipline as a database migration you can roll back, and teams that already do that for releases find it carries over pretty well. The ones who struggle are the ones who gave the agent their own login on day one.

Kerry Doyle 2 hours ago

had this last month lol. stopped an agent mid refactor and it left the branch half renamed, took me longer to untangle than doing it myself would have

Log in to join the discussion.