KubeCon NA 2026: kagent 1.0, llm-d 1.0, and VMs Still Run 53% of Apps

KubeCon + CloudNativeCon North America 2026 opened its main program in Salt Lake City today, 10 November, after a day of co-located events on 9 November. It is the first North American KubeCon with a dedicated AI Inference + Agentic track, and the project news arriving with it is about running agents and inference in production: kagent v1.0.0 was tagged on 9 November, llm-d v1.0.0 shipped on 20 October, and a CNCF-sponsored survey published on 9 November found that virtual machines still run 53% of production applications. I build a microVM platform, so the survey's isolation data was the part I read twice.

What happened at KubeCon North America 2026

The event runs 9 to 12 November at the Salt Palace Convention Center. CNCF executive director Jonathan Bryce set the theme when the schedule was published in August: "AI is quickly becoming one of the largest compute workloads the industry has ever seen." Monday's co-located events included Cloud Native AI + Inference Day, ArgoCon, BackstageCon and CiliumCon. The main program, which opened this morning, has three highlighted tracks: AI Inference + Agentic, Platform Engineering and Security. This article covers what was released or published around the opening; keynote announcements later in the week are not included.

Three releases and one report define the AI story so far:

• kagent 1.0. The CNCF Sandbox project for building and running AI agents on Kubernetes (accepted May 2025) tagged v1.0.0 on 9 November, and Solo.io, which created it, is promoting it at the show as "Now GA!". The release notes add an API v2 with AgentInstance lifecycle (suspend, resume, checkpoints and forks), an A2A gateway, sandbox agents with durable session state, harnesses for Claude Code and Codex, and RFC 8707 audience binding on token exchange.
• llm-d 1.0. The distributed inference project, a CNCF Sandbox project since March 2026, went 1.0 on 20 October. Its themes are multi-tenant quality of service with load shedding for bursty agent traffic, zero-downtime operations with a DisaggregatedSet that manages prefill and decode replicas as one unit, a v1 router configuration API, and agent-aware session affinity. It also deprecates building its own inference images in favour of upstream vLLM and others.
• Hyperlight for Linux apps. A CNCF member post from Microsoft on 9 November showed the Sandbox project Hyperlight running unmodified Linux programs inside hypervisor-isolated micro-VMs through hyperlight-unikraft, with snapshot restores of about 3 ms.
• The workloads survey. Linux Foundation Research, sponsored by CNCF and Broadcom, surveyed 492 practitioners in September 2026 about VMs, containers and agentic AI.

The numbers

If you want to check the release dates yourself, the GitHub API gives them directly:

# Requires the GitHub CLI (gh) and a logged-in session.
gh api repos/kagent-dev/kagent/releases/tags/v1.0.0 --jq '.published_at'   # 2026-11-09T18:43:27Z
gh api repos/llm-d/llm-d/releases/tags/v1.0.0 --jq '.published_at'         # 2026-10-20T02:30:10Z

Why it matters

Version 1.0 is a promise about APIs, and both projects made it in the same month. For llm-d that means the router config is v1 and disaggregated prefill and decode serving is something you can roll out and upgrade as a unit, which is the hard part I described in LLM inference serving in 2026. For kagent it means agents become Kubernetes objects with a lifecycle (create, suspend, checkpoint, fork) that platform teams can manage with the tools they already have.

The survey is the reality check. Two years of "Kubernetes is the operating system for AI" talk, and VMs still run more than half of production apps, with 71% of those answering expecting the share to stay in the same range next year. Agents are one reason: 47% of VM-expanding respondents cited security and isolation requirements, and 30% plan to use microVM sandboxing within a year, up from 20% today. Kubernetes is winning the control plane while the hypervisor keeps the isolation job. That matches what I argued in choosing isolation for agent workloads.

Disclosure: I run PandaStack, an Apache-2.0 Firecracker microVM cloud for agent code execution, so a survey saying microVM sandboxing is growing is convenient for me. Read the percentages, not my enthusiasm.

What sceptics say

The survey has limits its own authors flag. It is 492 respondents, more than one can represent the same organisation, Broadcom (which sells VMware) co-sponsored it, and the isolation figures are plans, not deployments. MicroVM runtimes as a deployment environment were cited by only 2% of respondents. A "planned within a year" jump from 20% to 30% is intent, which is cheap.

On the projects, 1.0 is a maturity claim made by maintainers. Both remain CNCF Sandbox projects, the lowest tier, and the kagent v1.0.0 tag came after four release candidates in ten days (rc1 on 30 October). The Gateway API Inference Extension that llm-d builds on is at v1.6.2, so the stack underneath is still moving. And agent identity has no settled standard in Kubernetes yet.

What it means for engineering and DevOps teams

• Pilot on 1.0, pin the versions. If you were waiting for llm-d or kagent to stabilise, now is a reasonable time to run a staging trial. Pin chart and image versions and read the deprecations first, especially llm-d's move to upstream inference images.
• Do not plan a VM exit for AI's sake. The data says mixed estates are normal. Put effort into shared identity, policy and observability across VMs and containers rather than a migration.
• Pick isolation per workload. Untrusted, agent-generated code belongs in a microVM or gVisor-class sandbox, not a plain container. Hyperlight's numbers show the startup cost argument against VMs is weakening.
• Give agents their own identities now. Even without a standard, run each agent under its own service account and scoped tokens. kagent's RFC 8707 audience binding is a small step in that direction.

FAQ

When and where is KubeCon North America 2026?

KubeCon + CloudNativeCon North America 2026 runs 9 to 12 November 2026 at the Salt Palace Convention Center in Salt Lake City, Utah. Co-located events were on 9 November and the main program runs from 10 to 12 November.

Is kagent 1.0 generally available?

kagent v1.0.0 was published on GitHub on 9 November 2026, and Solo.io is promoting it as generally available at KubeCon. The project itself remains at CNCF Sandbox maturity.

Do companies still run virtual machines alongside Kubernetes?

Yes. In a September 2026 survey of 492 practitioners by Linux Foundation Research, VMs averaged 53% of production applications, and 51% of respondents said at least half their production apps run as VMs. Only 8% of VM users manage them with KubeVirt.

Related: Kubernetes as the Agent Control Plane in 2026, Container vs microVM vs gVisor: Choosing Agent Isolation and LLM Inference in 2026: Prefill Is a Different Machine Now.