Bart Farrell: We know who you are, but you do need to tell us who you are, what's your role, and where do you work?
Abdel Sghiouar: Hi, everyone. My name is Abdel. I'm a developer advocate at Google Cloud, and I'm a co-chair of KubeCon North America.
Bart Farrell: Speaking of which, as one of the KubeCon program co-chairs, you've had a unique perspective on the technical conversations shaping this year's conference. What's one thing you think the Kubernetes community is getting fundamentally wrong about AI, and what are the consequences for engineers building real systems?
Abdel Sghiouar: I think that most people are just taking AI agents and sticking them inside pods and just calling it a day. And that's probably causing quite a lot of problems. One of them is a waste of compute resources. The second one is what I talked about in my keynote, which is the identity problem. An AI agent is fundamentally different than an app in the sense that it's unpredictable. So if you put it inside the container, inside the pod and rely on the service account to identify it, you are going to run into problems eventually, and you wouldn't actually know how to trace back the problem to the agent that calls it.
Bart Farrell: We're seeing projects like gVisor, Agent Substrate, OpenRL, and KubeRay moving towards CNCF governance. Are we building the foundations of an open AI infrastructure ecosystem? Or are we at risk of repeating the fragmentation and operational complexity that Kubernetes has struggled with for years?
Abdel Sghiouar: I think that both of these are true. Those components that have been donated to the CNCF are some of the components we actually need to turn Kubernetes into a viable platform for running containers, for running agents. But one of the things that a lot of people don't get, and I recommend you go check Agent Substrate's architecture. Agent Substrate relies on Kubernetes for the infrastructure part, but it has its own control plane and all the scheduling mechanisms for everything around agents. So I don't think that we are going to run into this fragmentation problem. I think what's going to happen in the future is that all these projects will leverage Kubernetes for the infrastructure automation. But then they will build abstraction layers on top of it, which serves the purpose that project is trying to achieve. This was a discussion I had with Tim Hockin. Some of these features will make their way eventually upstream to Kubernetes because they're just useful outside of AI workloads.
Bart Farrell: Your keynote raises the question of identity for AI agents. When an autonomous agent accesses a database, modifies infrastructure, or calls another service, whose identity and permissions should it inherit? What breaks in our existing security models? And how should platform teams start addressing it?
Abdel Sghiouar: That's a very good question. When we talk about identity for agents, we have to distinguish two different things. Either user identity, because the prompt was triggered by a user, and that prompt is what caused that particular agent to access a database or modify a service or call an API. In that case, we could use the user identity. So the agent will impersonate the user. Or if we're talking about autonomous agents, which act on their own or are triggered by workflow, in that case, we need to be able to give the agent its own identity. And that's where the SPIFFE and SPIRE come into the picture, where we can assign identities per instance, per session for each agent. And then with logs and traceability, we have a full end-to-end visibility into which agent assumed which identity and executed which actions. For platform engineers, that basically means leveraging existing tools, but just using them in different ways to manage identity for AI agents.
Bart Farrell: Looking three years ahead, what's one widely accepted assumption about Kubernetes and AI infrastructure that you think will turn out to be completely wrong? And what should engineers be experimenting with today to prepare for that possibility?
Abdel Sghiouar: That's a very good question. I think that isolation is going to be a big deal. Sandboxing your workloads in general, it doesn't matter if they are agents or not agents. It's something that people have to start experimenting with today. The funny thing about this is the technology has existed for a very long time. And we're in 2026 telling people to go look into it and use it. There are a lot of things that Kubernetes does today that has some limitations around agentic AI, like the churn rate and how fast etcd is and how fast the API server is. And these are things that are going to fundamentally break going forward if we keep deploying AI agents and we'll have to find new ways to deploy them. The second one, which was part of the keynote, is running virtual machines controlled by agents. Because virtual machines give you the sandboxing technology, but people still want to control them by Kubernetes. There is also this VM API, which was done by VMware, which is going to be a big deal in the future.