“Wikipedia recently talked about how the least retrieved content by humans was being frequently retrieved by agents. Because it was infrequently retrieved by humans, it wasn't in their caching system, and it was putting much more load on the system. They had built their system to handle human patterns they could predict, and the agents just broke that.”
The New Stack Podcast · Hands-On Engineering Podcasts · November 2026
Context
The CNCF's Jonathan Bryce uses Wikipedia's experience to show how agent traffic is bursty and unpredictable compared with human traffic. His point is that the question is no longer just how to run AI, but how to build systems that stand up to AI usage.
Transcript
…and you need to make sure that people's bank accounts have the right balances, or whether you're a frontier lab and you need the most flexible networking possible so you can fire up tens of millions of sandboxes and let people's agents run. There's some component in the CNCF to help you do that.
Okay, now you mentioned Kubevert. There's also Carmada, Kubeflow, Buildbacks, and I think one or two other things that are graduating. I want to move on to LLND really quickly, but in brief, which one of those do you think is the most important to raise for our friends today here on the show?
Carmada is an interesting one because I think that when, you know, we talked a little bit about some of the technology to enable AI workloads. Carmada can do that, but Karmada is also good for managing scalability and kind of making your clusters more resilient to unpredictable workloads. And that's really a signature of how agents treat our systems and our infrastructure. You know, Wikipedia recently talked about how they had this issue where the kind of like least retrieved content by humans was being frequently retrieved by agents. And because it was infrequently retrieved by humans, it wasn't in their caching system. And it was putting much, much more load on the system because they had built their system kind of handle human patterns that they could predict. And the agents just broke that. So Carmada is actually like a way to manage bursty and unpredictable load within a cluster or across clusters. So I think that's a really interesting one because it's actually not just like how do you run AI, but how do you build systems that can stand up to AI usage?
Yeah, I'm looking at the most viewed Wikipedia pages and I don't think agents are going to be into this. US Senate, YouTube, Facebook, US, and Donald Trump. I don't think those are deep in the weeds of what agents are looking for. That's a really interesting point. Now, the last time you heard…
Yeah. So it has actually just recently achieved its 1.0 launch. And the thing that I think the progression, which has been really rapid, because this is a project that's less than a year and a half old, maybe right about a year and a half old, it went from something that was built around a really…
Well, I'm really glad that you're out here single-handedly as a foundation saving Neo-Cloud economics. There's a lot of people invested in that. So I'm glad that we're figuring out. No, the memory crunch is a real problem around the world. And I recently bought a Mac Mini. And if you look at what…
More from The New Stack Podcast
