I want to run coding agents on a k3s cluster, start tasks from my phone, and keep their access limited. Would Kata microVMs be a sensible setup, and what else would need locking down?
How would you isolate coding agents on k3s and give them tasks from a phone?
Answers
Kata is a reasonable isolation layer for this. Each agent workload gets a guest kernel rather than sharing the host kernel like an ordinary container. That gives you a stronger boundary, although it doesn't make the agent's credentials or network access safe by itself.
Use throwaway pods for the agents and keep any state you need in Git or a persistent volume. If you want to run several agents at once, check the capacity of your nodes first. Separate microVMs have a resource cost.
Reply to this answer
I'd focus on what the agent can do after it starts. Give it read-only cluster permissions for debugging, with only the extra resources it actually needs to inspect. Exclude Secrets and use short-lived tokens.
For code changes, let it open pull requests but not merge them. Protect the deployment branch and have your deployment controller sync only that branch. That leaves a review step between an agent's proposed change and the cluster.
Reply to this answer
For phone access, look for an agent control option that connects outbound over HTTPS. That can avoid exposing an inbound SSH port or setting up a VPN just to hand off tasks. A default-deny egress policy is a good starting point, then allow the services the agent actually needs.
Reply to this answer
Another route is to use repository issues and pull requests as the mobile interface, with self-hosted runners handling the work. An orchestration service can map events to specific agents, prompts and toolsets, such as sending a labeled bug to a debugging agent or a review request to a review agent.
That suits repeatable workflows. For arbitrary tasks across different repositories or non-code work, a more general remote control interface may be simpler.
Reply to this answer
Add your answer
Answers are reviewed before publication.
q.481~velnaro:
Read-only access can still expose useful operational details, so the scope of those extra permissions is worth checking too.