Repository navigation
Evolving Kubernetes for Agentic Workloads: Rethinking the Pod and Node Boundary #126
Replies: 5 comments
|
I'm hoping this does not conflict with #117 so I can attend. |
|
I was about to propose a similar topic - perhaps we could cover both. In particular I want to talk about what we can do to optimize and enhance kubernetes on-node management for 2-layer systems. How can we make kube even better for hosting pods which themselves are systems inside. Examples:
|
|
@thockin All topics you listed above are the topics I want K8s and agentic ecosystem jointly address. Thanks! |
|
This topic has been selected and added to the schedule for 2:20 PM: https://events.linuxfoundation.org/kubecon-cloudnativecon-north-america/features-add-ons/maintainer-summit/?id=1318479#schedule Thank you for leading this discussion! |
|
For the two-level control model, could suspend/resume preserve an executor identity without preserving its authorization decision? An actor can sleep through a project/tool grant being revoked while the outer Pod and ServiceAccount remain unchanged. Our work on Sinaps Studio at SinapsData uses per-agent tool authorization at dispatch. Placing revalidation explicitly in the framework contract, with enough actor/resume identity to perform it, would keep that responsibility clear without requiring an etcd write per tool call. |
Uh oh!
There was an error while loading. Please reload this page.
Topic
Evolving Kubernetes for Agentic Workloads: Rethinking the Pod and Node Boundary
What projects are involved?
Kubernetes, Containerd, CRI-O, Kata, gVisor
Description
[Unconference] Evolving Kubernetes for Agentic Workloads: Rethinking the Pod and Node Boundary
Target Audience & Relevant Projects
Problem Statement
Autonomous AI agents represent the fastest-growing workload category in cloud infrastructure, yet their operational profile inverts almost every fundamental assumption behind Kubernetes container orchestration:
Because the Kubernetes Pod contract is fixed at creation time and every state change writes through etcd, emerging agentic frameworks (such as Agent Substrate, and many others) are increasingly routing around Kubernetes APIs on their hot paths, building custom out-of-tree control planes, running nested runtimes, and treating Kubernetes merely as an over-provisioned VM manager.
If Kubernetes does not evolve its process layer, the cloud-native ecosystem risks fracturing into proprietary sandbox fabrics.
Proposed Discussion Topic
This session will bring together maintainers across Kubernetes and the agentic ecosystem to align on how the core primitives should evolve:
Desired Outcomes
Session lead contact information
@dchen1107: Dawn Chen
@aojea: Antonio Ojea
All reactions