Ignite week always buries you in announcements. Strip the keynote gloss and one story is left: Azure is rebuilding itself around agents as the primary workload. Models, gateways, Kubernetes, and observability all changed this week with that assumption baked in.

What shipped

Area Announcement Status
Models GPT-6.2 Astra and GPT-6.2 Sol in Foundry GA
Governance Azure AI Gateway (standalone service) GA
Kubernetes Managed GPU node pools, CVM node pools on Azure Container Linux GA
Kubernetes AKS AI Manager for Foundry, Agent Sandbox extension, Pod Snapshot Preview
Agents Foundry continuous improvement, Foundry IQ knowledge graphs New

GPT-6.2 and the cost-per-task argument

GPT-6.2 Astra and GPT-6.2 Sol are generally available in Foundry. Astra is the reasoning-heavy model; Sol targets production agents, coding, and high-frequency work.

The interesting claim isn't capability. Microsoft says Astra scored higher average accuracy than GPT-6 Astra with substantially fewer output tokens in Foundry testing, at the same per-token price. That's a vendor benchmark, so verify it on your own evals. But the framing is right: stop comparing price per token and start measuring cost per task.

The other detail worth your attention is 12-hour cache writes on Standard deployments, alongside the 30-minute default. If you run agents that go quiet overnight, long-lived system prompts and tool definitions can now stay cached at discounted read rates. This is the same discipline I argued for in context engineering: the cheapest token is the one you don't re-send.

If you're still deciding how to call these models, my breakdown of choosing the right AI SDK still applies.

Azure AI Gateway is the boring announcement that matters most

Azure AI Gateway is now GA. It's the former AI Gateway tier of API Management, spun out as its own fully managed product with an AI-only roadmap.

What you get:

  • Models from Foundry, Azure OpenAI, OpenAI, Anthropic, Amazon Bedrock, Google Vertex AI, or any OpenAI-compatible endpoint, called by an alias you define, so swapping models needs no app change.
  • Routing by priority, weight, or condition, with failover.
  • MCP server federation, plus turning OpenAPI specs and APIs into MCP tools.
  • Entra ID auth, policy enforcement, and usage telemetry on every request.

Pricing is freemium: 10,000 free requests per gateway per month, with paid billing starting January 1, 2027. It launches in 17 regions.

My take: every team running more than one agent will end up building this layer by hand, badly. Credentials scattered across apps, no central cost view, and no kill switch. A managed gateway is the right default. The multi-provider support is the tell: Microsoft knows you're not single-vendor, and it's choosing to be the control plane instead of fighting that.

AKS gets serious about agent workloads

The AKS Ignite roundup is the densest of the week.

Foundry on your own clusters

AKS AI Manager for Foundry (limited preview) lets you deploy open-weight models and Foundry hosted agents onto AKS clusters you already run. Developers stay in the Foundry workflow and the same azd flow; platform teams keep their networking, identity, policy, and GPU capacity. For regulated shops, this is the version of Foundry that gets past the security review.

Sandboxes for model-generated code

Three previews attack the same problem, running untrusted agent-generated code without making users wait:

  • Agent Sandbox extension, built on the open-source Kubernetes Agent Sandbox project, with pre-warmed pools and either Kata microVM isolation or a serverless runtime on Azure Container Instances.
  • Pod Snapshot saves a pod's memory and process state to Blob Storage and restores it on demand.
  • A new kata isolation runtime class that starts pods from a pre-initialized VM template with lower memory overhead.

Per-session isolation is the correct security model for code-executing agents. I covered the broader shift in agentic DevOps as the next evolution of shift-left: the guardrails have to live in the platform, not in the prompt.

GPUs and confidential nodes

  • Managed GPU node pools are GA: AKS manages the driver, device plugin, and metrics exporter, with metrics flowing into Azure Monitor.
  • NVIDIA GB200 and GB300 sizes and a mixed MIG strategy are in preview.
  • Confidential VM node pools with Azure Container Linux are GA on Intel TDX and AMD SEV-SNP, along with a CIS Level 1 benchmark for the OS.

Managed GPU pools remove an entire category of "why is the driver mismatched" tickets. That alone justifies the upgrade.

Foundry learns to improve agents

Microsoft also announced continuous improvement in Foundry: feed traces and user input back into the system, let Foundry find and test improvements, see the trade-offs across quality, speed, and cost, and then you decide what ships. Microsoft also claims its model router can save up to 65% versus sending everything to a frontier model. Treat that as a ceiling, not a forecast. There's also Foundry IQ with knowledge graphs for agent grounding.

The human-approves-the-change step is what I like. Self-modifying agents with no review gate are an incident waiting to happen.

What it means

  1. Evaluate before you migrate models. Run GPT-6.2 Sol against your current production agent and compare cost per completed task, not token price.
  2. Put a gateway in front of your agents now. The free tier makes a pilot cheap. Mind the January 1, 2027 billing date.
  3. Most of the AKS agent features are previews. Pilot sandboxes in non-production; adopt the GA GPU and CVM features today.
  4. Watch observability. There's also a What's new in Azure Monitor at Ignite 2026 post worth skimming if you run agent telemetry.

Bottom line

Last year Azure sold you models. This week it sold you the whole operating environment for agents: identity, gateway, sandboxes, GPUs, and a feedback loop. The pattern is clear: Microsoft wants to be the governed runtime underneath whatever model you choose. If that holds, the competitive ground moves from "who has the best model" to "who makes agents safest to operate", and that's a DevOps problem, which is good news for those of us who've been solving it for years.