Azure Weekly

Issue #561 - 3rd May 2026

Highlights this week, starting with AI: DeepSeek V4 Flash and V4 Pro in Microsoft Foundry: Microsoft has added DeepSeek V4 Flash to its Foundry model catalogue (with V4 Pro coming soon), giving teams a single platform to mix a low-latency model for high-volume work with a deeper-reasoning model for complex agentic tasks. Bring Your Own Model to Foundry Agent Service Is Now Generally Available: BYOM in Foundry Agent Service has reached general availability, letting enterprises connect prompt agents to any Chat Completions-compatible model routed through Azure API Management or a third-party AI gateway. Memory in Foundry Agent Service (Public Preview): Foundry Agent Service has gained a managed long-term memory store in public preview that automatically extracts, consolidates and retrieves user preferences and conversation context so agents stay coherent across sessions. Microsoft Agent Framework 1.0 (GA): Microsoft Agent Framework has reached version 1.0 GA for both .NET and Python, providing a production-ready, open-source SDK that unifies Semantic Kernel and AutoGen into a single stack for single- and multi-agent workflows. Automate agent evaluation with the Evaluation APIs: Copilot Studio's Agent Evaluation is now exposed through Power Platform REST APIs and a certified connector, so teams can wire test-set runs into CI/CD pipelines, Power Automate flows or custom tools as quality gates.

Fabric April 2026 Feature Summary: April's Microsoft Fabric round-up brings tabbed multitasking to GA, AI auto-descriptions for semantic models, deeper VS Code integration and notebook retry policies, cross-workspace MLflow logging, ALTER TABLE in explicit transactions, and several new Real-Time Intelligence capabilities including streaming mirrored database change feeds into Eventstreams. Multi-layer Caching with the Decorator Pattern: endjin's Jonathan George walks through layering an in-process IMemoryCache over an Azure Blob Storage cache in front of a Databricks SQL Serverless source, using the Decorator pattern (and a per-key semaphore) to defeat cold-start latency and the thundering-herd problem for an analytical web API.

VS Code Curbs Token Use Ahead of Copilot's Controversial Usage-Based Billing Switch: Two days after GitHub announced Copilot's switch to usage-based billing on 1 June 2026, Microsoft shipped VS Code 1.118 with prompt caching tweaks, deferred tool loading and smaller purpose-built models for search and execution, all aimed at cutting the tokens (and therefore the bill) per agent turn. Microsoft-OpenAI rewrite opens the door for Anthropic and Google: OpenAI and Microsoft have amended their partnership so OpenAI can serve its products across any cloud (including AWS Bedrock) while Microsoft is freed to build its own models and lean more heavily on rivals like Anthropic and Google, with revenue-share payments to Microsoft continuing through 2030.

Finally, Hardening OpenClaw on AKS: Mitigating Container Escapes with Kata microVM Isolation: A step-by-step walkthrough of running the OpenClaw autonomous AI agent inside Kata Container microVMs on Azure Kubernetes Service, swapping the shared-kernel container boundary for a hypervisor boundary so a successful container escape lands inside a guest VM rather than on the host.