The honest practitioner's take on agentic AI on Kubernetes | BRK222
Lachlan Evenson walks through how agentic AI workloads differ from “normal” services and what it takes to run them reliably on Kubernetes in production.
Overview
This Microsoft Build 2026 breakout focuses on production considerations for agentic AI on Kubernetes, including:
- Why agentic systems tend to be stateful, bursty, and multi-step, and why they often span more than a single cluster.
- How Kubernetes can be used for different AI workload types:
- Inference
- Training
- Agentic systems / agent orchestration
- Scheduling challenges for large AI workloads, including the need for gang scheduling.
- Layering AI platforms on top of Kubernetes, including an architecture segment covering AnyScale Runtime integration with Azure.
- Managed and purpose-built tooling options for running AI workloads, plus open-source inference at scale.
- A demo segment that includes Azure KARS (described as a secure sandbox) using workload identity.
- A closing segment on AKS Claw and building agentic workloads on top of Kubernetes.
Session resource
Chapters (from the video description)
- 0:00 - Agenda Overview and Enthusiasm for Agentic Systems
- 00:09:41 - AI Workload Types: Inference, Training, and Agentic Systems Explained
- 00:12:36 - Challenges of scheduling massive AI workloads and need for gang scheduling
- 00:14:12 - Transition to layering AI platforms on top of Kubernetes
- 00:25:29 - Architecture Overview: Integration of AnyScale Runtime with Azure
- 00:30:06 - Launching and Managing AnyScale Workspaces and Clusters
- 00:35:05 - Conclusion: Agentic Orchestration and Future of Autonomous Agents
- 00:38:19 - Launching demo of Azure KARS: secure sandbox using workload identity
- 00:47:36 - Final segment on AKS Claw and building agentic workloads atop Kubernetes