Bob Killen of the CNCF on agents and Kubernetes

Who Bob Killen is

Bob Killen is Director of Technical Program Management at the Cloud Native Computing Foundation. The role, in Bob’s own description, is split between the foundation’s projects and its end users, and much of the work is bringing the two together: the people who build Kubernetes and the 230 or so projects around it, and the companies that run them in production.

That position gives an unusual view of the ecosystem. The end users say what is hurting; the projects say what they are building. When the two do not line up, Bob’s team is the one that notices first.

The conversation

Bob sat down with the show at Open Source Summit Europe in Prague on 8 October 2026, together with Daniel Krook, the CNCF’s Senior Director of Developer Experience, in a session arranged by the Linux Foundation’s press team. The full conversation runs about fifteen minutes: from how a project moves from sandbox to graduated, to what AI agents are doing to the platform those projects run on.

Twenty requests for every human one

The part of the conversation that belongs most clearly to Bob is about load. For as long as there have been web applications, the shape of traffic has followed a person: one request, wait for the answer, make the next one. Agents break that pattern. One human action can now fan out into twenty agent requests, each of which needs somewhere to run.

With a pod in Kubernetes, the guarantee of spinning one out is about one second. And now with the agents, you’re looking for sub-200-millisecond response time.

— Bob Killen

A second is fine for a microservice that scales over minutes. It is far too slow for an agent that expects an answer in a fifth of that. Kubernetes itself, Bob says plainly, is not the best fit for that kind of workload as it stands, and the gap is being closed by new projects that keep warm capacity ready so agents can start almost instantly, at very large numbers.

Identity, isolation and the hardware underneath

The rest of the conversation, shared between Bob and Daniel, follows the same load into the layers below. The first is identity: once an agent acts on someone’s behalf, the platform has to know which agent it is and which person started the request. Projects such as SPIFFE and SPIRE, built for machine and workload identity, are where that work starts, alongside newer projects that declare what an agent may and may not do.

The second is hardware. Much of the interesting work in the CNCF right now is the software layer above scarce accelerators: slicing GPUs and other devices so that each agent request gets what it actually needs, rather than a whole card it will barely use. Right-sizing the request on the way in matters as much as having the hardware at all.

The advice for people starting out

Asked what a newcomer should do with an ecosystem this large, the answer was practical. Get hands-on and pick something, whether it is popular or simply interesting. Build a small local inference setup, try a couple of options, and learn the underlying principles before reaching for the complex hosted version. The common pitfall is jumping into Kubernetes without real container knowledge first.

Where to go next

The written interview with Bob and Daniel on lucaberton.com, linked at the top of this page, has more detail, and so does Daniel Krook’s own article. For the CNCF as a whole, and everyone the show has interviewed there, see the CNCF on Udienza. The rest of the Prague week is in the show’s write-up of Open Source Summit Europe 2026.