Microsoft Developer Video Summaries

This video walks through the Microsoft Copilot harness, showing how to build custom agents that can incorporate additional skills, tools, and sub-agents. It covers the setup process, configuration options, and the built-in capabilities that let you scale without managing the underlying infrastructure.

■ Agent Setup and Configuration
- Creating a new agent and enabling the harness option
- Defining a name, skills, and MCP servers for the agent
- Specifying the harness type and operating instructions

■ Built-in Harness Tools and Capabilities
- Using sub-agents and helper agents for tasks such as queries
- Fetching pages from the web
- Creating, editing, and deleting files in a local workspace via an Azure YAML file

■ Scaling and Deployment
- Tool calling within the agent definition
- Scaling without worrying about the infrastructure in between
- Using the harness within Microsoft Foundry

This content is aimed at developers and engineers who want to build custom agents using Microsoft's tooling. After watching, you will understand how to configure an agent with skills and tools, what built-in capabilities the harness provides, and where to start building in Microsoft Foundry.

This session demonstrates how to build scalable prompt agents using Microsoft Copilot Harness within Microsoft Foundry. A cupcake ordering agent is used as a practical example to show how adding skills, sub-agents, and MCP servers can handle increasingly complex scenarios without requiring custom infrastructure or container hosting.

■ Core Concepts
- How Microsoft Copilot Harness scales prompt agents by adding skills, sub-agents, and built-in tools
- Types of built-in harness tools, including parallel task handling, web page fetching, and workspace file management
- The role of MCP servers in extending agent capabilities

■ Live Demo: Cupcake Ordering Agent
- Building an agent with instructions, tools, and multiple skills
- Adding a single-order skill, a bulk-order skill using CSV input, and a party-planner skill
- Reviewing and confirming orders through a dashboard

■ Agent Configuration with Azure YAML
- Defining skills, MCP servers, and tools in a YAML file
- Setting up the harness type as GitHub Copilot (preview) within Microsoft Foundry

This session is geared toward developers and AI engineers who want to build and scale prompt agents without managing infrastructure. Viewers will learn how to extend agents with modular skills and MCP servers, and gain a practical understanding of agent configuration in Microsoft Foundry.

MicrosoftのPatrick LeBlancを迎え、Fabric Database Hubに搭載されたDatabase Agentsの機能と今後のロードマップを紹介します。データベース管理を「問題が発生してから対応する」モードから、エージェントが問題を自動検知し通知する継続的・証拠ベースの運用へシフトさせるというコンセプトが中心です。実際にブロッキングの検知からSSMSやGitHub Copilotを使った調査までのデモを含みます。

■ Database Agentsの概要とシフト
- リアクティブなDBA業務から、問題を先回りで検知・通知するプロアクティブな運用への移行
- 現時点では問題の検知・検出に対応し、調査・推奨はSSMSやCopilotで継続
- 11月第1週をめどに全世界で順次提供予定

■ デモ: 設定とブロッキング検知
- 管理ポータルとDatabase Hubでの有効化手順（SQL対応、PostgresとCosmosも順次追加予定）
- ブロッキング、クエリパフォーマンス、インデックスチューニング、リソース圧迫の監視シナリオ
- Agent Previewタブでの問題表示と、リードブロッカーとなるクエリの特定

■ 調査フローと今後のビジョン
- SSMSとGitHub Copilotを用いた調査・修正のワークフロー
- エージェントにメモリを持たせ、対話の蓄積を通じて信頼を獲得し自動化へ段階的に移行する構想
- Kustoベースのスキルを使った対話的なデータベース状態確認のプロトタイプ

DBA、データベース開発者、DBA業務に関わるエンジニア、Microsoft Fabricの利用者や関心がある方向けの内容です。Database Agentsの設定方法から実際の問題検知・調査フロー、将来の方向性までを一貫して把握できます。説明欄に前回のDatabase Hub回のリンクも掲載されています。

Microsoft Azure's Michael McKenzie explains how IoT, industrial AI, and physical AI work together to move enterprises from telemetry collection to real-world action. The talk covers the spectrum of customer maturity, from gathering visibility to running learning loops on machines and processes.

■ IoT and the Path to Physical AI
- How IoT connected machines, meters, vehicles, and smart products for fleet visibility
- Why visibility alone does not improve business outcomes
- IoT for telemetry, industrial AI for reasoning, physical AI for action
- How physical AI learns with each execution and builds learning loops
■ Azure Services for Connectivity and Operations
- Azure IoT Hub: connecting millions of devices globally with secure maintenance and updates
- Azure IoT Operations: data curation and inference in operational technology environments
- Azure Device Registry: applying IT policy, role-based access control, and certificate management to OT devices
- Using Fabric and Foundry to gather, analyze, and generate business outcomes
- The shift from connecting more things to making things and processes smarter

IT and OT leaders, product managers, and engineers evaluating Microsoft's IoT and AI platform will get a clear framework for where each service fits and how the pieces complement one another.

Christoph Berlin, vice president of engineering at Cloud AI at Microsoft, explains how agentic AI can be deployed across factories and plants at scale. He addresses the practical constraints of running agents across 300 sites with latency, real-time, and resiliency requirements, and outlines a blueprint for a repeatable and composable agent architecture.

■ Deployment Reality and Constraints
- Why not everything can be moved to the cloud and why running 300 separate on-premise instances is impractical
- The case for a hybrid model with one agent framework that can run anywhere
■ Recommended Blueprint
- Common data contracts to avoid building everything bespoke
- Guardrails for secure policy handling across environments
- Agent runtime options: cloud, edge, and fully disconnected scenarios
- Managing one agent implementation across many sites through a single interface
■ Key Learnings
- How agentic AI accelerates business processes without replacing them
- How the right architecture makes agentic AI repeatable rather than bespoke

This talk is aimed at engineering leaders, manufacturing IT professionals, and AI practitioners designing agent deployments at scale. It helps viewers understand the architectural principles and trade-offs involved in scaling agentic AI across industrial sites. Further reading is available at microsoft.com/ai/manufacturing.

Azure Front Doorグループでプリンシパルソフトウェアエンジニアとして働くKarthik Gopinathan氏に、CDNがインターネットの安全性と速度を支える仕組みや、近年急増しているDDoS攻撃への取り組みを語ってもらいます。

■ DDoS攻撃の仕組みとAIによる影響
- 分散型サービス拒否攻撃（DDoS）の基本概念と、増幅攻撃やプロトコル脆弱性の紹介
- AIの登場が攻撃手法の発見・武器化をいかに加速しているか
■ ホリスティックな防御への転換
- 攻撃ベクトルごとの個別対応から、システムリソースの監視とトラフィック相関分析による全包括的防御への移行
■ 白板を使った具体的な分析手法の解説
- 確率的データ構造によるTop Kトラッキング、IPグループ分類、脅威インテリジェントの活用
- クライアント署名のプロファイリングでボットとブラウザを区別し、良性トラフィックを守る方法
- 数秒単位のスパイクにリアルタイムで対応するためのシステム設計
■ 開発の裏側
- 大規模なDDoS攻撃をきっかけに、複数組織が協力して約1年かけて構築した経緯

セキュリティエンジニアやクラウドネットワークに関わる方に向けて、DDoS対策の設計思想と実装アプローチを具体的な図解とともに理解できる内容です。

A weekly roundup of Microsoft developer news, covering the latest releases and updates that matter for building and shipping applications. This edition focuses on .NET 11 RC2, Python on Windows on Arm, browser automation in Microsoft Foundry, and new Azure capabilities.

■ .NET 11 RC2
- First build in the release cycle with a go-live support license for production use
- Spans runtime, SDK, libraries, MSBuild, NuGet, F#, ASP.NET Core, .NET MAUI, and Windows Forms
- Supported by Visual Studio 2026 Insiders and Visual Studio Code with C# Dev Kit
- How teams can validate compatibility, performance, deployment pipelines, and breaking changes before the stable release

■ Python on Windows on Arm
- Python 3.15 selects the native arm64 runtime by default as a tier-2 CPython platform
- Coverage of PyPI project wheel availability and common workflows such as NumPy, Django, FastAPI, Jupyter, and Pytest
- Role of native CI runners for package maintainers

■ Microsoft Foundry and Azure
- General availability of browser automation using managed Playwright workspaces
- Features including toolbox integration, isolated sessions, live view, human takeover, tracing, and Entra ID / Azure RBAC support
- New Azure tools for operations, offline apps, and hardware-rooted security

Aimed at developers and engineering teams working with the Microsoft stack, this roundup helps you decide which updates to plan for and what to validate in your own projects. Watching lets you quickly grasp what changed, what is now supported for production, and which areas to explore further in the official announcements.

ガードレイルを設定せずにエージェントを公開すると、ユーザーの悪意あるプロンプトや脱線した質問にエージェントが乗せられてしまうリスクがあります。本エピソードでは、Microsoft AIエージェント向けのガードレイル（プリデファインドガードレイルとカスタムガードレイル）を用いて、エージェントをプロンプトインジェクションやPII漏えいから守る手順を、カップケーキ注文エージェントを作りながら実演します。

■ 保護すべき脅威の整理
- ジェイルブレイク、プロンプトインジェクション（直接・間接）への対策
- 暴力・自傷・性的コンテンツなどのコンテンツセーフティ
- 顧客の氏名・住所などのPII漏えいの防止
- タスク逸脱（本来の業務から逸れた応答）の抑制

■ ガードレイルの作成と割り当て
- Microsoft定義ガードレイルと独自ガードレイルの使い分け
- エージェント単位でのガードレイルの割り当てと保存

■ ガードレイルの効果検証テスト
- 指示の書き換えによる攻撃の試行
- 業務外の依頼（ジョークなど）への対応確認
- ビジネスルールを破るトリックの試み
- ボーチャルコードや顧客ID・支払い情報の聞き出し試行
- 安全ポリシー違反となる発言へのブロック確認

エージェントを業務に組み込もうとしている開発者や、Microsoft AIエージェントのセキュリティ設計に関わる方に向けて、ガードレイル設定から実際の攻撃シナリオでのテストまでの一連の流れが確認できます。視聴後は、自作エージェントにもガードレイルを設定し、同様のテストで挙動を検証するところまで実践できます。

An examination of how AI can support factory operations and business processes rather than replace human workers, focusing on practical uses in investigation and evidence gathering.

■ AI in factories
- Framing AI as assistance for people and processes
- Addressing the question of whether AI replaces humans

■ Root cause analysis example
- Correlating data across multiple systems
- Gathering evidence and coordinating with subject-matter experts
- Using AI to reach evidence faster and prioritize investigation steps

Useful for viewers in manufacturing, operations, and business-process roles who want a practical way to frame AI adoption. It offers a clearer perspective on AI as an accelerator for existing workflows and human decision-making.

This Data Exposed episode features Ikra from the SQL product management team discussing agent-led growth for Azure SQL and how changing coding tools affect developer workflows. It introduces the Azure SQL developer hub and Microsoft SQL agent skills, then demonstrates building and moving a task-tracking app with GitHub Copilot in VS Code. The demo shows local development with an Azure SQL DB container and migration to Azure SQL Database in the cloud.

■ Developer ecosystem and Azure SQL strategy
- Coding tools moving beyond suggestions toward personal agents
- The question of where application data lives and Azure SQL's role

■ Azure SQL developer hub and agent skills
- Getting-started guidance, sample prompts, sample apps, and coding tool choices
- Microsoft SQL agent skills that help agents understand SQL
- Plugin support for GitHub Copilot and VS Code, plus other coding tools

■ Demo: local to cloud workflow
- Building a task-tracking app with GitHub Copilot and the MSSQL extension
- Using the Azure SQL DB local container, currently in private preview
- Migrating schema and app to Azure SQL Database cloud with Microsoft Entra
- Free-tier Azure SQL Database as an option

■ Production readiness and feedback
- Scaling from a side project to production with Azure SQL
- Feedback on the GitHub repo and upcoming updates

This is suited to developers and data professionals exploring AI-assisted application development with SQL. Viewers can learn the overall workflow presented and are encouraged to visit the Azure SQL developer hub, try the sample prompt, join the container preview, and provide feedback.

This episode starts a series on building a production-style AI travel planning application with Azure Cosmos DB, beginning from a blank local folder and ending with a working app without an Azure subscription. It focuses on setting up a local Cosmos DB environment and using AI-assisted development tools in VS Code to design a data model, generate application code, create queries, and test locally.

■ Local development setup
- Starting the Azure Cosmos DB emulator with a container command
- Running a local Cosmos DB environment without provisioning cloud resources

■ AI-assisted workflow
- Using a short PRD and a Barcelona long-weekend scenario as the application example
- Applying GitHub Copilot, the Azure Cosmos DB Agent Kit, and the MCP Toolkit for data modeling and code generation
- Generating and refining NoSQL and vector queries from natural language
- Seeding the local emulator with travelers, trips, destinations, hotels, restaurants, and flights

■ Scope and next steps
- The episode covers local setup, data model design, code generation, query creation, and local testing
- It does not cover cloud deployment or vector, full-text, and hybrid search, which are introduced as the next episode

Developers who are new to Azure Cosmos DB or want to understand AI-assisted application development can use this episode to follow a local setup and workflow. Viewers can apply the described steps to move from a specification to a testable local application and prepare for the next episode on intelligent search over operational data.

This episode of Code with Azure Cosmos DB explains a lifecycle for operating production AI applications with confidence, focusing on observing, evaluating, and optimizing agent behavior. It continues a series that builds a travel planner application and shows how operational data, vector search, and agent memory reside in Azure Cosmos DB.

■ Observability
- Traces capture tool calls, routing decisions, and memory retrievals
- LangSmith, Microsoft Agent Framework, and Microsoft Foundry provide visibility into agent behavior
- Open telemetry allows traces to flow into existing observability stacks

■ Evaluation
- LangSmith supports evaluation of routing decisions, tool calls, and end-to-end performance
- LLM-as-a-judge patterns, regression testing, and CI/CD evaluations help catch issues before production
- Azure Cosmos DB-specific evaluation verifies cross-session preference recall

■ Optimization
- RU optimization through indexing strategy, partitioning model, and query patterns
- Retrieval tuning for vector and hybrid search
- Memory summarization and compression
- LLM token reduction through tighter prompts, smaller context windows, and fewer model calls

Viewers building AI applications on Azure gain a framework for understanding agent behavior, catching regressions, and improving production efficiency. The episode is suitable for developers and product teams interested in observability, evaluation, and optimization practices for stateful agents. Viewers can apply the observe-evaluate-optimize approach to their own AI application lifecycle.

This episode continues the Code with Azure Cosmos DB series by adding persistent memory to a multi-agent travel planner. It explains why AI applications need memory across sessions and how Agent Memory Toolkit APIs store and retrieve memory in Azure Cosmos DB. The discussion covers stateful agents, shared memory across specialized agents, and preference updates over time.

■ Memory foundations for agents
- The default problem of applications starting without preferences, history, or context
- Four memory types: short-term, long-term, episodic, and semantic
- Using Agent Memory Toolkit APIs for consistent memory storage and retrieval in Azure Cosmos DB

■ Multi-agent travel planner example
- Research, hotel, and restaurant agents working from a shared user picture
- Agent hand-off through a single source of truth rather than direct messages
- Cross-session continuity and explicit preference corrections in plain English

■ Architecture takeaways and next steps
- Treating memory as a foundational part of multi-agent application design
- A preview of production concerns such as observability, result verification, and failure detection

This is aimed at developers and architects building AI agents and multi-agent systems with Azure Cosmos DB. Viewers gain a conceptual framework for persistent shared memory and an example of how specialized agents can coordinate across sessions. It can help inform how to design memory-aware agent applications and prepare for production considerations.

This episode of Code with Azure Cosmos DB builds on the travel planning application started in episode one, showing how hybrid search combines vector similarity with full-text precision and how a single database can power both retrieval and booking workflows. The presenters are Marko Hotti, Senior Product Marketing Manager for Azure Cosmos DB, and Patty Chow, Product Manager at Azure Cosmos DB.

■ Search capabilities
- Vector search, full-text search, and hybrid search, each serving a different purpose
- Semantic reranking within hybrid search, demonstrated with a Barcelona hotel search
- How exact name matches and semantic results are handled together

■ Data architecture
- Keeping operational data and vector embeddings in one database instead of separate systems
- One SDK, one security model, and one database to operate
- The distinction between retrieval queries and transactional reads and writes

■ Booking workflow
- Moving from property search to date selection, availability checks, and reservations
- How inventory, bookings, and preferences are handled under the same system

■ What comes next
- Preview of the following episode: specialized agents and shared memory using the Agent Memory Toolkit

Developer teams building production AI applications that must retrieve live operational data and complete user tasks will find this useful. Viewers can expect to understand when to apply each search type, why keeping vectors and app data together affects retrieval accuracy, and how search and transactional operations fit into one system.