For those of you running AI agents, what’s actually painful right now?

Curious to hear from people who are using AI agents in real workplace settings, not just personal projects or demos , What’s been the most annoying part of getting them into production or through an internal or customer security review? I keep hearing about things like audit trails, prompt injection, runaway loops, and not really knowing what an agent can touch once it has access to real tools But I’d rather hear from people actually dealing with it. If you’re using agents at work, or trying to get them there, what’s the biggest friction point right now? Have you had to put together any kind of evidence or logs for a security team? And what would make you feel more comfortable letting an agent interact with real systems or customers? Even small war stories are welcome,I’m trying to understand what’s genuinely hard in practice versus what just sounds scary in theory P.S I am working on an open source project and would love to have other developer helping me or contributing to this cause 😄 DM me if you want to work on this too