The CNCF on Udienza: Kubernetes meets agents

About the Cloud Native Computing Foundation

The Cloud Native Computing Foundation is the Linux Foundation home of Kubernetes and about 230 other open source projects. Around 40 of them are graduated, the most mature level, around 40 are incubating, and a long tail sits in sandbox; the Technical Oversight Committee, elected from the governing board, the maintainers and the end users, decides who moves up.

Two cities, one argument

The show has met the CNCF twice this year, and the two conversations land in the same place.

In Yokohama, at KubeCon Japan, the foundation’s CTO, Chris Aniszczyk, put it in one line: agents are kind of the new microservices, a new type of workload. They are non-deterministic, often bound to GPUs, sometimes to custom accelerators, but they are still distributed systems that fail and have to be scaled, and that is what ten years of Kubernetes and the projects around it were built for.

In Prague, at Open Source Summit Europe, Bob Killen and Daniel Krook filled in what that means in practice. One human action now fans out into some twenty agent requests. A pod takes about a second to start, and agents want an answer in under 200 milliseconds. The ecosystem’s response is the one the CNCF always has: established projects such as gVisor extended to isolate agents, and new sandbox projects for the problems that are genuinely new.

The common thread for a platform team: nobody at the foundation is telling you to rebuild for agents. They are telling you which layers will be under the most pressure, the cold start, isolation, identity and the scarce hardware underneath, and where to watch for the projects that relieve it.