- Vishakha Sadhwani
- Posts
- KubeCon NA 2026
Agents, sandboxes, new CNCF projects and new certs ~ everything from Salt Lake City that's worth your attention.
Hi Inner Circle!
KubeCon + CloudNativeCon North America just wrapped up in Salt Lake City (Nov 9–12).
Through the keynotes, the sessions and the announcements ~ here’s all you need to know (obviously from my lens)
If I had to sum up the week in one line:
Kubernetes is no longer just running containers ~ it's becoming the control plane for AI agents.
That sounds like a big claim.
But almost every keynote came back to the same few questions:
→ How do you run thousands of agents without paying for idle compute?
→ How do you isolate code that an agent wrote?
→ And who is that agent, when it calls your database?
Let's go through it, one bucket at a time.
1. Agents are now a Kubernetes workload
This was the biggest theme of the week.
As you know, an agent is a model + a harness.
→ The model does the thinking. It usually runs somewhere else ~ behind an API, or on a GPU inference server.
→ The harness is the code around it. It holds the conversation, calls the tools, and runs the loop until the job is done.
The harness is the part that runs in your cluster. On Kubernetes, that's a pod.
Here's the catch.
Most of the time, the harness is just waiting ~ for the model to reply, a tool to finish, or a human to answer.
So if every agent gets its own pod, that pod sits idle while it waits.
And you're still paying for it.
Agent Substrate (Google → CNCF Sandbox)
Google donated Agent Substrate to the CNCF, along with gVisor and OpenRL.
Instead of one pod per agent, many agents share a small pool of warm workers. Idle agents get paused, and resumed when they're needed.
→ Why not make each agent a Kubernetes object?
The API server isn't built for millions of fast-changing objects. So Agent Substrate keeps agent state outside it, and lets Kubernetes manage the workers.
It's now at 1.0, and available as a managed feature on GKE.
gVisor → isolation
Agent Substrate → agent runtime
OpenRL → reinforcement learning on Kubernetes
OpenShell (NVIDIA → CNCF Sandbox)
NVIDIA contributed OpenShell, its open source runtime to secure and govern autonomous agents.
NVIDIA and Google are also working on integrating OpenShell with Agent Substrate.
So one handles where the agent runs, and the other handles what it's allowed to do.
Agents are treated as untrusted
This came up again and again in the keynotes.
Agents are nondeterministic. So the guidance was to contain them out-of-band, before they touch your network or a production database.
The tools that kept getting named:
Kata Containers and gVisor for isolation
SPIFFE/SPIRE for short-lived identity
Envoy-based gateways to enforce policy
Who is the agent?
One keynote title summed this up well: "The Identity Crisis: Who is the AI Agent in Your Cluster?"
→ When an agent calls a tool or a database, what identity does it present?
Identity is where authorization and audit start. So expect a lot more on this in the next year.
Operating at "agent speed"
On Day 2, Jonathan Bryce shared that agents now generate around 70% of tracked inference traffic.
Agents don't follow human patterns. They crawl, retry, and fan out.
The takeaway was that platforms need to be more self-healing, with guardrails enforced by policy ~ rather than manual tickets and fixed capacity planning.
2. Projects that moved up
The CNCF portfolio is now at 231 projects.
A few moves worth knowing:
NATS graduated. It's the messaging system used by 2,000+ companies, across cloud and edge.
Graduations called out on stage: OpenTelemetry, KEDA, Kyverno, Cloud Native Buildpacks, CubeFS and KubeVirt.
10 projects moved from Sandbox to Incubating, including CloudNativePG, Microcks, Confidential Containers, Fluid, KubeRay and Open Cluster Management.
And a couple of newer ones:
OpenRL ~ abstracts reinforcement learning building blocks, and schedules training and sampling across GPU pools using Dynamic Resource Allocation (DRA).
VM API (Broadcom) ~ brings Kubernetes-style declarative management to virtual machines.
→ Why are VMs showing up at KubeCon?
Because isolation. 34% of organizations said they're expanding VMs specifically for AI and agent workloads.
3. New certifications
This one is for everyone who asks me which cert to do next.
Available now ~ Certified Kubernetes Network Engineer (CKNE)
It covers Kubernetes networking, security and troubleshooting for production environments. It’s also hands-on! So you can see why it’s super cool to add to your profile!
Also thia exam assumes CKA-level admin knowledge.
One beta tester also mentioned AI networking topics like InferencePool and traffic management for LLM workloads.
Coming in 2027
Kubernetes AI Associate (KAIA) ~ entry level
Certified Kubernetes Inference Engineer (CKIE) ~ hands-on
These were announced, but you can't register yet.
So if you're preparing, CKNE is the one you can look into today.
4. Networking and platform updates
A quick run through the releases that touch day-to-day infra work:
Istio Ambient multi-network is GA ~ basically a sidecar-less service mesh, plus support for migrating from sidecars to Ambient is available.
AKS (Microsoft) added managed DRA, Fleet placement controls and upgrade health gates.
NVIDIA DSX AI Factory got three new open source components, including NVIDIA Kubernetes Engine (NVKE) and Snapshot for faster recovery of GPU inference.
Apple container can now create clusters ~ one command gives you an upstream Kubernetes cluster built with kubeadm, where each node is its own Linux VM on your Mac.
Uber is open-sourcing Michelangelo, its Kubernetes-native ML platform.
→ Which one should you try first?
If you're running a service mesh, the Ambient migration path is worth a look.
If you're on a Mac and want a local multi-node cluster, Apple container clusters might be the easiest one to play with this weekend.
5. What real teams are running
Two stories I'd point you to, because they're about scale and not demos.
Meta ~ GitOps for AI infrastructure
Meta won the CNCF End User Case Study contest.
Around 15 engineers run ~100 Kubernetes clusters for AI research and inference. They moved from Terraform on cron to GitOps with Argo CD and Kargo.
→ What changed? Deployments went from hours to minutes.
CVS Health ~ one logical cluster
CVS built its platform in about ten months, across three hyperscalers and private data centers.
Dozens of Kubernetes clusters show up as one logical cluster, using an automated placement engine, OpenCost, and multi-cluster service meshes with cross-region failover.
Governance is built in, so both developers and AI agents stay compliant by default.
Both stories point the same way ~ fewer people managing more clusters, with GitOps and policy doing the heavy lifting.
6. Ops tooling is going agentic
Almost every observability and ops vendor showed some form of AI agent this year.
Jaeger previewed AI-assisted troubleshooting and trace visualization for GenAI workloads.
Testkube added agent pipelines over MCP ~ agents run tests, read logs, find the root cause and open a fix PR.
Solo.io made its enterprise Agent Substrate and kagent generally available.
Mirantis launched a self-hosted platform for running governed AI agents on Kubernetes.
→ Are teams letting these agents act on their own?
Mostly not yet.
Testkube shared that most customers keep agents read-only, or keep a human approving the change.
That feels like the right place to start.
7. Numbers worth knowing
21 million cloud native developers ~ up 36% in a year
4x growth in inference traffic across providers in about eight months, with network traffic growing faster than compute
~70% of tracked inference traffic now comes from agents
34% of organizations are expanding VMs for AI and agent workloads
75% of cloud native developers use AI coding tools, vs 55% of non-cloud native developers
The one I keep thinking about is the network one.
We usually talk about GPUs as the bottleneck. These numbers suggest networking might be catching up.
What should you learn next?
If you're a DevOps or infra engineer, here's what I'd put on the list after this week:
Sandboxing ~ how gVisor and Kata Containers isolate a workload
Workload identity ~ SPIFFE/SPIRE, and how a workload proves who it is
DRA (Dynamic Resource Allocation) ~ how GPUs get scheduled now
Istio Ambient ~ service mesh without sidecars
KubeVirt ~ running VMs and containers side by side
GitOps at fleet scale ~ Argo CD and Kargo across many clusters
You don't need all six at once.
Pick the one closest to what you run today, and go one level deeper.
Final thoughts
None of these tools are brand new ideas.
Isolation, identity, scheduling and GitOps have been around for years.
What's changing is the workload sitting on top of them.
Agents are just another workload ~ but a noisy, unpredictable one that needs stronger boundaries.
The fundamentals still matter. They just matter more now.
See you in the next one,
Vish
• • •