Introduction
John Savill published two deep dives this week: sovereignty considerations for Azure and NAT Gateway v2. The sovereignty one is worth watching if you’re dealing with compliance or data residency requirements. He walks through the full spectrum from standard public cloud sovereignty capabilities all the way to Azure Local disconnected, covering legal ownership, control plane, identity, and data. The NAT Gateway v2 video covers zone redundancy, IPv6, higher throughput, and flow log support.
On to the updates. The standout for me this week is instant snapshot access for Premium SSD v2 and Ultra disks. It’s one of those quality-of-life gaps that’s been quietly annoying people, and now it’s fixed. The AKS MCP Server preview is also interesting if you’re building AI agents that need to interact with your clusters.
Premium SSD v2 and Ultra disk instant snapshot access
If you’ve used Premium SSD v2 or Ultra disks, you’ve probably hit this: you create a snapshot, and then you wait. Unlike standard HDD, SSD, and Premium SSD where snapshots are immediately usable, these two disk types needed backend data hydration before you could do anything with them.
That delay is gone. You can now opt in to instant access, and the snapshot is usable the moment you create it. Spin up new instances, restore, whatever you need. The data hydration still happens, but it runs transparently in the background without affecting your performance.
This is especially useful if you’re snapshotting before major OS or application changes and want a quick rollback path, or if you’re cloning production into test environments and don’t want to sit around.
AKS node autoprovision: encryption at host and local DNS
Two additions to AKS node autoprovisioning, which uses Karpenter under the hood. Quick reminder: Karpenter doesn’t just autoscale nodes. It looks at what pods need to be scheduled and picks the optimal node size and SKU, which tends to save you money by fitting workloads more efficiently.
First, encryption at host is now supported. This encrypts the temporary storage on the node, any caching, the ephemeral OS disk (which is very common in AKS), and the path between managed disks and the host. All of that is now encrypted without you having to change your workload configuration.
Second, local DNS support. This puts a per-host DNS cache and forwarding layer on each node. Your pods get lower DNS latency because of the local cache, and transient DNS issues upstream won’t break your pods because cached records are still served. If you’ve ever had pods fail because of a brief DNS hiccup, you know why this matters.
Node autoprovisioning is also now available in Azure Government and private cloud environments.
AKS MCP Server (preview)
AKS has an MCP server now, available on GitHub.
If you haven’t worked with MCP yet: an MCP server exposes its capabilities, knowledge, and tools in a standard way so that AI applications and agents can discover and use them without the developer having to explain everything manually. The agent talks to the MCP server, the server tells it what it can do, and the agent works with it.
With the AKS MCP server, your AI agents can perform CRUD operations on AKS resources, get cluster details, and even manage multi-cluster scenarios through fleet support. If you’re building agents that need to interact with Kubernetes infrastructure, this is the integration point.
AKS Agentic CLI (preview)
Related to the MCP server, the AKS Agentic CLI lets you interact with your AKS clusters using natural language. Run az aks agent and just ask questions: “what’s wrong with my cluster?” and the AI will guide you to specific issues.
The agent component can run either as a local container on your machine or as a pod on the AKS cluster itself. Seems most useful for troubleshooting, where you know something is off but aren’t sure where to start looking.
Azure Functions Flex Consumption tools
The Flex Consumption SKU for Azure Functions has new troubleshooting tools available. Quick context: Flex Consumption is the Linux-based serverless option that combines pay-for-what-you-use billing and scale-to-zero with features you’d normally associate with higher tiers: private networking, instance memory size selection, always-ready instances, and fast large-scale-up.
The new tools help diagnose quota issues, related constraints, and general deployment and configuration problems. If you’ve been running Flex Consumption and hit issues that were hard to debug, these should help.
Azure Functions .NET 10 support
Azure Functions now supports .NET 10 in preview, on both Windows and Linux using the isolated worker model.
The isolated model runs your function in a separate worker process with its own .NET version, independent of the host runtime. This gives you version flexibility and avoids conflicts. If you want to start testing your functions against .NET 10, the path is there.
Claude Sonnet 4.6 on Azure
Claude Sonnet 4.6 from Anthropic is now generally available across Azure AI Foundry, GitHub Copilot, Azure Databricks AI model serving, and other Microsoft services.
Sonnet is Anthropic’s mid-tier model: faster and cheaper than Opus, but still capable. It comes with a 1 million token context window, which means you can feed it very large codebases, operational data, or multi-turn conversation histories. Output caps at 128K tokens.
It also supports computer use, where the model can interact with terminal sessions and regular user interfaces to perform actions on surfaces that don’t have an API or MCP server. And it adaptively decides whether reasoning is required for a given task, so it doesn’t burn tokens on thinking when a straightforward answer will do.
Agent Monitoring Dashboard (preview)
The Agent Monitoring Dashboard is in preview as part of Microsoft Foundry. I think this one is more useful than it sounds at first.
It gives you both operational and evaluation metrics for your AI agents. On the operational side: token usage (spot inefficiencies), latency, and success rates. Those are standard observability metrics.
Where it gets interesting is evaluations. With non-deterministic models, you need to track whether your agent is staying fit for purpose over time. Is the quality of its output degrading? Is it drifting from its intended behavior? You set up evaluations to measure this, and the dashboard surfaces the results.
It also shows red teaming results, surfacing security and safety risks. If you’re running agents in production and don’t have a way to monitor whether they’re still doing what you built them to do, this is worth a look.
Final thoughts
Instant snapshot access for Premium SSD v2 and Ultra disks is the kind of update that doesn’t make headlines but removes real friction. If you’ve been working around the hydration delay, turn on instant access and move on.
The AKS updates keep piling up. Between the MCP server, the Agentic CLI, encryption at host for node autoprovision, and local DNS, there’s a lot happening in the AKS space. The MCP server in particular feels like where things are heading: AI agents that can operate your infrastructure through standardized interfaces.
And the Agent Monitoring Dashboard deserves attention. We’re at the point where organizations are deploying agents, but the tooling to verify they’re still working correctly over time has been lagging behind. Evaluations, token usage, safety metrics, all in one place. That’s the kind of thing you’ll wish you had set up before you needed it.
Sources
- John Savill, “Azure Update - 20th February 2026,” YouTube, https://www.youtube.com/watch?v=4EVWtkjKTkI
- John Savill, “Looking at Sovereignty Requirements in Azure”, YouTube, https://www.youtube.com/watch?v=Q6Anv-8a-I0
- John Savill, “NAT Gateway v2 Overview”, YouTube, https://www.youtube.com/watch?v=I9HPGJUjq2Q