I've been diving into various ways to use agents as interaction models on Kubernetes, and I want to hear from others about their experiences. In particular, I'm curious if anyone has implemented a human-in-the-loop system for delegating tasks to these agents. How did you set it up, and what's your approach to safely letting humans delegate tasks within that system? Also, do you prefer using a centralized agent/microcontroller server, or do you find using local setups to be more effective? Personally, I lean towards a local model approach since it seems like a practical tool where humans remain accountable for their actions on the cluster. However, I find it challenging to delegate specific tasks to a model easily. What do you all think?
3 Answers
Have you guys checked out Stakpak? It's open source and vendor neutral, which might give you more flexibility with AI in your Kubernetes setup. You can find it on GitHub.
I regularly use AI to generate manifest boilerplate and help troubleshoot issues in Kubernetes. One tip: set up a read-only context for the AI when running commands outside a lab environment. This way, it can only read data without making changes. Recently, I asked about a Cilium Loadbalancer issue and found some useful options through AI. Just remember: never let it run write commands without supervision! Keeping your brain engaged is key!
Thanks for sharing your approach! Setting permissions seems key, but what if you don’t have those elevated permissions? Are there simpler ways to restrict AI's actions?
Honestly, I think the idea of delegating tasks to agents can be risky. Giving an agent too much access without fully trusting it is dangerous. I’d avoid that if possible! But, if you have systems in place to limit its abilities, maybe it’s not as bad. I’m not a fan of agents myself, but if they’re unavoidable, ensuring they're tightly controlled makes sense for safety.
I get that! I think any tool can be a liability if you don't handle it right. Maybe using very strict permissions to limit what the agent can do could help manage that risk.
Great advice! How do you set up that read-only context? Is it just about creating a restricted ServiceAccount, or are there other methods you would recommend?