At Microsoft Ignite 2026 this week, we're introducing AKS AI Manager for Foundry, which lets you deploy open-weight models and Foundry hosted agents from Microsoft Foundry onto Azure Kubernetes Service (AKS) clusters you already run. Developers stay in the Foundry workflow they know, and platform teams keep their own networking, identity, policy, and GPU capacity. Alongside it, AKS is shipping updates for agent sandboxing, GPU operations, confidential computing, node lifecycle, and multi-cluster management. Here's what's new and where to start.
For the wider set of Azure AI infrastructure news, including Agent Grid and Azure Cobalt 200 VMs, read the Azure AI infrastructure announcement from Ignite.
Foundry models and agents on AKS clusters you govern
Many organizations want the Foundry developer experience but need AI workloads to run on infrastructure their platform team already secures and pays for. AKS AI Manager for Foundry brings the two together, bringing intelligent optimization to AI workloads, GPU and VM performance, topology, and real-time capacity allocation to decide where workloads land and how the infrastructure is configured.
Deploy open-weight models from Foundry to AKS (limited preview)
You pick an open-weight model in Foundry and deploy it to your own AKS cluster with the security policies and governance already in place. Your team chooses GPUs and regions based on cost, latency, and data residency while Foundry handles the setup, so developers keep one workflow from first test through production. At Ignite, BRK420 walks through the whole path, from selecting a model and a runtime tuning profile to serving it on GPU node pools with workload identity, policy, and built-in observability.
Run Foundry hosted agents on your own AKS clusters (limited preview)
With this deployment option, Foundry runs the agent runtime on your AKS cluster, and developers keep building, deploying, and calling agents through Foundry with the same APIs, Azure Developer CLI (azd) flow, and endpoint they use today. Each session runs in its own pod inside a lightweight VM through AKS pod sandboxing, which keeps sessions isolated from one another. Foundry manages the agent lifecycle and its integration components, while your team operates the cluster, node pools, networking, container registry, and storage.
This option fits teams that already run AKS as a platform, or that need agent workloads to stay inside an infrastructure, network, or compliance boundary they manage. Sign up for the hosted agents on AKS preview, and read what's new for hosted agents in Foundry Agent Service for the rest of this week's agent updates, including network egress controls and memory snapshots.
Isolated sandboxes for agent-generated code
Agents that execute model-generated code need an isolated environment for each session, but starting a fresh runtime every time leaves users waiting, and three previews work on both sides of that problem. The Agent Sandbox extension for AKS, built on the open-source Kubernetes Agent Sandbox project, handles sandbox creation, lifecycle, and pre-warmed pools, with two runtime options: Kata microVM isolation on your cluster's own nodes or a serverless runtime enabled by Azure Container Instances (ACI). The Pod Snapshot cluster extension saves a supported pod's memory and process state to Azure Blob Storage and restores it on demand, so a sandbox picks up where it left off instead of initializing again. Outside of agent workloads, pod sandboxing also gets a new kata isolation runtime class (preview) that starts pods from a pre-initialized VM template with lower per-pod memory overhead, so any sandboxed workload gets VM-level isolation with faster starts.
Less GPU setup and better GPU utilization
Managed GPU node pools are now generally available (GA). On each enabled node pool, AKS manages the GPU driver, Kubernetes device plugin, and metrics exporter, and GPU metrics flow into Azure Monitor. Support for NVIDIA GB200 and GB300 VM sizes is in preview on Ubuntu and Azure Linux node pools, with AKS managing the GPU software stack for Grace CPUs and Blackwell-family GPUs. If your GPUs are sitting underused, the mixed multi-instance GPU (MIG) strategy (preview) splits each supported GPU into differently sized instances, so workloads with different requirements share one GPU.
Confidential computing and hardened nodes
Confidential VM (CVM) node pools with Azure Container Linux are now GA on Intel Trust Domain Extensions (TDX) and AMD Secure Encrypted Virtualization-Secure Nested Paging (SEV-SNP) VM sizes, pairing hardware-based protection for data in use with an immutable, container-optimized host operating system. Sensitive workloads move over without refactoring, and the Center for Internet Security (CIS) Level 1 Benchmark for Azure Container Linux, also GA, gives security teams a published baseline to assess nodes against.
In preview, Ubuntu 26.04 CVM node pools add Intel TDX support and encrypt the operating system (OS) disk by default from first boot, with no configuration changes. Node Hardening (preview) reserves resources for the kubelet, container runtime, and host services and adds soft eviction thresholds, so busy nodes stay responsive and pods shut down gracefully under memory or disk pressure.
Nodes that scale out and recover with less intervention
Node auto-provisioning (NAP) creates right-sized nodes in response to pending pods, and this release brings it to more workloads. Trusted Launch, Capacity Reservation Groups, instance-level public IP addresses, OS disk type selection, and AKS Machine API integration are now GA, while Windows Server nodes and node image pinning are in preview. NAP also adds Capacity Buffers, which keep a declared amount of spare capacity ready for bursty or latency-sensitive workloads in place of placeholder pods.
On the recovery side, pod eviction on freeze is GA: before a scheduled VM freeze, AKS cordons the node and evicts only the pods you've labeled, respecting Pod Disruption Budgets. In preview, node autorepair can bring up a replacement node before removing an unhealthy one, and Container Network Interface (CNI) autorepair detects supported pod startup failures and starts recovery on its own. To confirm your applications hold up through events like these, Azure Chaos Studio workspaces and scenarios are GA, with pod-level faults for AKS clusters and a report for every run.
Zone-redundant egress and simpler fleet operations
New clusters using the managed network address translation (NAT) gateway outbound type now get a managed StandardV2 NAT gateway (GA) where StandardV2 is available, adding zone redundancy, up to 100 Gbps of throughput, IPv4 and IPv6 addresses, and flow logs. Existing clusters keep the NAT gateway they have, so moving to StandardV2 stays your call.
For teams running many clusters, Azure Kubernetes Fleet Manager brings several capabilities to GA. Cross-cluster networking provides a managed Cilium multi-cluster network with global service discovery, now including member clusters on overlay networks. The security patch auto-upgrade channel applies weekly Linux node security updates, and scheduled start gates and label-based update groups decide when an update stage begins and which clusters it reaches. Multi-cluster identity binding management, in preview, lets you define an identity binding once and distribute it across member clusters.
AKS desktop and AI-assisted troubleshooting
AKS desktop picks up three previews: support for AKS on bare metal clusters without opening an inbound firewall port, an AI Assistant that detects GitHub Copilot and brings in AKS Skills and AKS MCP (Model Context Protocol) for cluster-aware troubleshooting, and Headlamp plugins with a Plugin Catalog for GitOps, security, and autoscaling workflows. For air-gapped environments, AKS Investigator (preview) runs read-only, AI-assisted incident investigations locally on AKS on Azure Local disconnected operations.
Catch up on KubeCon North America
If you missed KubeCon + CloudNativeCon North America 2026 last week, Brendan Burns covered the upstream and AKS updates that sit alongside these launches, including managed dynamic resource allocation (DRA) for GPUs and GA placement controls in Fleet Manager. Read What's new in Kubernetes, open source, and AKS for the full rundown.
Watch the AKS sessions on demand
These breakout sessions are recorded and will be available on demand after they air.
- BRK420: Deploy and run Microsoft Foundry models and agents on AKS. Take a model or agent from Foundry to a governed deployment on your own cluster.
- BRK422: Run open-weight models, post-training, and agents on AKS. GPU capacity, model serving, and distributed post-training with Ray, with Harvey on stage.
- BRK421: Operate a global AKS fleet across regions, edge, and sovereign clouds. Keep distributed clusters consistent with Fleet Manager, with JetBlue on stage.
Where to start
If you want Foundry models or agents on infrastructure your team already governs, sign up for the hosted agents on AKS preview and watch BRK420 (live streamed and recorded). If you're running GPU workloads today, try managed GPU node pools on a test cluster first, and if you're building agents that execute code, start with the Agent Sandbox extension for AKS. We'd like to hear what you're trying and what's still getting in your way, so leave a comment below.