@SnoopJ
@audAsta [AMP]
@tante
@glyph
The thing about agents, from what I understand in talking to vendors about using them, is that to use them correctly you have to build very detailed and specific playbooks for them to "follow".
In practice, it seems like most people just think you can Claude your way to success with vibes and vaguery.
They seem to think having an agent eliminates the hard part: defining your process in clear language. In truth, it's more important because an agent won't have the "common sense" to not delete and recreate your production database at 4:30 on a Friday before a three day weekend. Or just delete it.
This is not even including the identity and access boundaries you need. Like, we are having deep discussions about an agentic solution that would just read help desk tickets and make suggestions to the help desk personnel. We have to consider all the ways prompt injection could abuse its access. And when the agentic AI is telling people what to do, that's a prime target for social engineering. They want it to be able to reboot servers. That's a denial of service attack waiting to happen.
An outside vendor we've spent lots of money on is trying to sell us a multi-agent system that management is already in love with and we have to educate them on the almost unfathomable risk it would create. How are they forgetting everything they've ever learned about risk modeling, threats, fraud, attack surfaces, least privilege, etc. These are not stupid people, but they are acting like wide-eyed children just because it has the word "AI" attached to it. They should be more skeptical, not less.