How do you actually stop an agent before it does something destructive?
retained text
I see a lot of posts about agents going wrong after the fact. Wrong account touched, a delete that shouldn't have run, a tool call that burned way more budget than it should have. What I can't find much on is stopping it before it happens. Right now all I really have is prompt instructions telling the agent what not to do, plus manual review for anything touching production. That works, but its kind of tiring . A system prompt is just so basic, not real enforcement. how are you devssecure your agents ( whcih primarily run my saas) not just telling it what not to do. is there any repo or software to catch prompt injection before it reaches a tool call. And do you have a spend cap that actually stops execution, not just alerts you after the fact. if anyone has every experienced anyhting like this, how did you guys fix it or did u find it?
document fetched 2026-09-18 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wiu6wi/how_do_you_actually_stop_an_agent_before_it_does