At KubeCon + CloudNativeCon North America 2026 in Salt Lake City, VMblog stopped by the Kubex booth to talk with Andrew Hillier, the company’s Co-founder and CTO. He explains how they are taking Kubernetes and AI infrastructure optimization beyond chatbots, toward agents that do real, human-approved work in a closed loop.

What is Kubex?

Kubex provides resource optimization for Kubernetes and AI infrastructure. That means auto-sizing and autoscaling containers, nodes and GPUs, and increasingly auto-tuning vLLM so AI models run efficiently and quickly. Andrew describes it as a full-stack focus, from Kubernetes right into AI.

What sets Kubex apart

• A production-proven automation controller. Andrew says optimization shouldn’t be manual, because it should reduce toil rather than add to it. Many large customers run the automation in production, where costs and risks go down.

• A new agentic harness. Many vendors offer a chatbot or an MCP layer. Kubex goes further by giving agents (its own, or ones like Claude, Devin or Cursor) the normalized workload models, policy models, analytics and audit trails they need to do productive work.

Live demo: from agent proposal to closed-loop optimization Andrew walks through the agentic harness on screen:

• MCP-based tools expose nearly all product features.
• Domain-specific skills cover bin packing, node pre-warming, PDB optimization and HPA optimization.
• Agent “tasks” run in the background, scanning for opportunities and raising a short, prioritized list of proposals instead of overwhelming teams.
• Each proposal has a human-readable rationale and a machine-readable Kubernetes resource that can be executed. In the demo, an agent recommends adding a memory-optimized node type to a Karpenter configuration.
• Engineers can approve, edit or ask the agent to revise the proposal. Only after approval does it pass to Kubex’s deterministic automation layer, so agents never touch automation directly.
• Everything is also available through MCP, so proposals can flow into a GitOps-style merge request process.

AI infrastructure and GPU optimization

• GPU fractioning. Kubex analyzes models running on whole GPUs and, where a half or quarter would do, automates fractioning through the NVIDIA KAI Scheduler. Andrew says this can cut millions of dollars in wasted GPU spend by bringing requests in line with utilization.
• vLLM autotuner (coming soon). Announced at the show, it runs locally in your cluster without the full Kubex product. It tunes runtime parameters such as KV cache, batching and sequencing to match traffic and prompt patterns, and it works alongside auto-fractioning. Early access is available by scanning the QR code at the booth.

Learn more

Visit Kubex’s website (documentation and free trial available) – https://kubex.ai Catch Kubex at KubeCon + CloudNativeCon Europe in Barcelona this spring

Back to: KubeCon + CloudNativeCon 2026