Azure Weekly

Issue 577

30th August 2026

Highlights this week include:

🤖 AI

  • Understanding Harnesses: The Concept That Changes How You Build Agents in Copilot Studio The GitHub Copilot harness in Copilot Studio introduces three distinct runtime layers-Chat, Standard, and GitHub Copilot-to allow developers to choose between model-agnostic goal-directed reasoning, explicit conversational flow construction, or traditional scripted agents, marking a significant evolution from the previous singular approach.
  • Building Production-Ready AI Agents in Microsoft Foundry: 10 Lessons Learned Building production-ready AI agents in Microsoft Foundry requires focusing on grounding, observability, resiliency, governance, and continuous monitoring to address issues like API timeouts, outdated content, tool failures, and unexpected token usage.
  • Building Enterprise RAG on Azure: GPT-4o + Azure AI Search + Azure DevOps — End to End Building Enterprise RAG on Azure combines GPT-4o with Azure AI Search’s hybrid search and Azure DevOps for seamless integration, ensuring precise, secure, and scalable generative AI applications.
  • Enabling A2A endpoint and Agent Card for a Hosted Agent With Microsoft Foundry’s hosted agents, developers can quickly set up an A2A endpoint and Agent Card, enabling other agents to discover and communicate with their containerized service via the standardized Agent-to-Agent protocol, while keeping configuration separate in control and data planes.
  • Reducing Token Usage Reducing Token Usage explores techniques like token management, context pruning, efficient history handling, lightweight summarization, TOON compression, and server-side computation to minimize token consumption in AI applications using the Microsoft Agent Framework.
  • Benchmark Retrieval Patterns for Copilot Studio and Foundry IQ The summary offers a reproducible framework to objectively compare retrieval patterns in Copilot Studio and Foundry IQ by separating latency, billing units, and quality signals.
  • Consume Dataverse MCP server in Microsoft Foundry-Bringing business data to AI Agents The Dataverse MCP Server allows AI agents to securely access and interact with Dynamics 365 and Power Platform data in real time, reducing custom integration needs to under 30 minutes.
  • How to evaluate LLMs before production Evaluating language models for production requires assessing real-world inputs, ambiguous labels, missing context, and distribution mismatches-issues not captured by clean benchmarks alone.
  • Deploying to Foundry Deploy the app using Docker, push to Azure Container Registry, register with Foundry, poll for active status, and invoke via its endpoint.
  • AI overconfidence risks The experiment assesses how various AI models impact code quality, architecture decision consistency, and associated risks in an inner source tool handling architectural choices.
  • Copilot Studio | Adding Dataverse Knowledge Sources with the Web API (the "UI-Only" Rule That Isn't) This article explains how to add Dataverse knowledge sources to Copilot Studio agents programmatically using the Web API, despite UI limitations, noting that these methods are undocumented and may change.
  • The Economics of Agent Optimization: Four ways to lower the cost Microsoft Foundry introduces four runtime levers to lower AI outcome costs by optimizing model selection, caching, prompt efficiency, and resource usage based on task complexity.
  • Grok 4.6 comes to Microsoft Foundry Models: Built for long-horizon reasoning and complex workflows Grok 4.6 for Microsoft Foundry Models is engineered for long-horizon reasoning and complex workflows, excelling across coding, engineering, office productivity, and technical problem solving with strong performance on benchmarks like Terminal-Bench 3.0 and 3DCodeBench.
  • The Dark Side of Autonomous Agents - How to Stop Them The presentation highlights how autonomous AI agents can become malicious execution vectors through manipulation of external file specifications, stressing the necessity of a zero-trust architecture with Policy Enforcement Points for security.
  • Adding a Fallback Model to Hermes with Microsoft Foundry Adding a fallback model with Microsoft Foundry to Hermes ensures redundancy and enterprise controls, allowing seamless failover when Bedrock is unavailable, all at pay-as-you-go token costs.
  • Evaluating a Ternary Protein Language Model Architecture for Rapid Antibody Sequence Generation The peleke-1bit architecture, using a ternary BitNet model with 26 tokens and 33.6 million parameters, achieves comparable antibody design performance to larger models while running efficiently on CPUs with minimal latency impact.
  • Cohere Parse v5 is now available in Microsoft Foundry Cohere Parse v5 integrates with Microsoft Foundry to convert enterprise documents into structured Markdown while preserving layout and supporting nine languages at up to 36 pages per second on an H100 node.
  • Introducing Cohere Parse V5 Cohere Parse V5 transforms enterprise documents into structured Markdown, handling tables, forms, diagrams, and images across nine languages at 4.5 pages per second per GPU on Azure Foundry.
  • Copilot Cowork browser use – First Look Copilot Cowork's new feature enables AI to perform website tasks directly in Edge with your sign-in, following existing security policies, but needs admin activation and is active only when you're present.
  • Azure Machine Learning AutoML for FinTech Startups: From Idea to Production Azure Machine Learning AutoML accelerates FinTech startups' path from concept to production by automatically training and assessing various ML pipelines for classification, regression, and forecasting, allowing swift creation of dependable models with minimal data science resources.
  • MCP vs. Tools vs. Skills: What AI Architects Need to Know Understanding the distinctions between tools, MCP, and skills clarifies how AI architects should design reliable secure maintainable low-latency cost-efficient systems.
  • Beyond Tokens: Rethinking AI Economics with Microsoft Foundry Microsoft Foundry redefines AI economics by measuring value through outcomes instead of token consumption.

🔎 Analytics

🖥️ Compute

🚢 Containers

🗄️ Databases

🛠️ Developer tools

🔩 DevOps

  • Your PC Is a Build Agent: Self-Hosted GitHub Actions in Docker Desktop Docker Desktop lets you run a self-hosted GitHub Actions runner locally via a simple Dockerfile and entrypoint setup, enabling seamless CI/CD workflows on your PC without cloud dependencies.
  • Copilot Code Reviews for Azure Repos (public preview) GitHub Copilot now provides onboarding controls at organization, project, and repository levels, supports Managed DevOps Pools, allows customizable instructions for team standards, triggers automatic reviews via branch policies, and offers enhanced billing visibility with Azure Cost Management tags.
  • GitHub Copilot app for Beginners: Automate Dependabot pull request triage GitHub Copilot’s new app automations simplify Dependabot pull request triage by grouping updates by risk, verifying CI status, and delivering a concise summary daily.
  • Azure DevOps in the GitHub Copilot App The Azure DevOps plugin for GitHub Copilot lets developers view, open, edit, and complete pull requests and work items directly within the Copilot interface without switching applications.
  • OpenClaw went viral. Meet the maintainers building and securing it. OpenClaw's swift growth from a personal project to an open-source success story with 388k stars showcases effective handling of massive pull requests, inclusive contributor engagement including non-developers assisted by AI, and adaptive trust models amid software supply chain challenges.
  • Copilot Studio | Real ALM for Agents (and the One Click That Breaks Your Pipeline) Copilot Studio's ALM path fails critically due to missing agent deployment support, reliance on undocumented "standard harness," and lack of automated solution import without authentication, making it unsuitable for real-world production.

🧬 Hybrid + multicloud

🎭 Identity

🔌 Integration

💡 Internet of Things

  • Reviving the Tessel 2 - The First Rung Reviving the Tessel 2 project aims to upgrade from OpenWrt 15.05 to 17.01 to test if Tessel-specific configurations and bridge software remain compatible with current upstream releases.

🎓 Learning and Certifications

  • What's New in Microsoft EDU - Back to School August 2026 Microsoft EDU launches Teach with Remember Settings for unified class management, multilingual AI content generation, and upcoming Unit Plans integrating standards, pacing, and assessments.
  • AI-901: Model Deployment Options and Configuration Parameters AI-901 focuses on selecting appropriate deployment options in Microsoft Foundry and understanding key inference parameters like temperature, max tokens, top-p, penalties, and system prompts to shape model output behavior for exam scenarios.

⚖️ Management and Governance

🌐 Networking

🔐 Security

📦 Storage

🔗 Web

Get Azure Weekly in your inbox every Sunday

Sign up to receive the email every week and keep on top of all the announcements.

Don't forget to confirm your subscription. We have a double opt-in sign up process to prevent spam.

We will never provide your details to any third parties. We don't spam.

Azure Weekly Logo