CHAPTERS
- 0:17 – 1:18
Session goals: build and deploy Claude-powered agents in Microsoft Foundry
Marlene opens by outlining a hands-on session: deploy a Claude model in Microsoft Foundry, connect it to an agent, and then equip that agent with tools. The talk is positioned as a practical workshop rather than a conceptual overview.
- •Workshop format: build along live rather than only theory
- •End-to-end outcome: model deployment → agent integration → tool enablement
- •Speakers and support staff introduced for in-room help
- 1:18 – 3:20
Why the industry is moving from chatbots to agentic systems
The presentation frames the shift from single-turn chat to agents that can plan, reason, and act over time. It highlights the new technical and operational challenges this shift creates for real-world deployments.
- •Agents require multi-step reasoning and long-context handling
- •Enterprise needs: reliability, observability, and security
- •Agents must connect to tools, external systems, and data sources
- •Model improvements alone aren’t enough—systems/orchestration matter
- 3:20 – 4:51
What Microsoft Foundry provides for building agents at scale
Marlene explains Microsoft Foundry as a unified platform for building AI applications and agents using familiar developer tools. She outlines Foundry’s core components and how it supports integration-heavy, enterprise-ready agent development.
- •Foundry integrates with GitHub, VS Code/Visual Studio, Copilot Studio
- •Core platform pieces: Models, Agent Service, Tools/Integrations, ML services
- •Large integration surface: 1,400+ connectors and MCP tools
- •Enterprise features: governance, observability, security; Entra ID integration
- 4:51 – 6:51
Why use Claude inside Foundry: reasoning, orchestration, and faster productionization
The talk motivates the Claude + Foundry combination by focusing on agent-quality reasoning and operational readiness. It emphasizes moving from prototype to production via built-in evaluation, monitoring, and enterprise controls.
- •Claude strengths: planning, long context, agent reasoning (e.g., Sonnet/Opus)
- •Foundry adds agent creation, orchestration, and tool access beyond the raw model
- •Out-of-the-box compliance/security and enterprise integrations
- •Shorter dev-to-prod cycle via evaluation and observability tooling
- 6:51 – 9:24
Workshop setup: launch the Skillable lab and navigate to the hands-on guide
Attendees are guided to the workshop link and instructed to launch a pre-provisioned environment. The session is self-paced-friendly, with step-by-step guidance available and support staff on hand.
- •Navigate to aka.ms/claude workshop and click Launch
- •Skillable provides a virtual machine environment
- •Self-paced option: follow ‘Next’ steps independently
- •Green “auto-fill” boxes reduce manual typing and errors
- 9:24 – 10:24
Scenario briefing: Sparkles cupcake shop needs an agent to handle high demand
The lab’s narrative is introduced: Sparkles has too many customers and flavors for staff to manage manually. The goal is to build an agent that can help manage ordering and customer flow.
- •Sparkles is overwhelmed by incoming cupcake orders
- •Agent will assist with customer interactions and ordering
- •Scenario sets up the need for tools/data access (inventory, flavors, vouchers)
- 10:24 – 12:56
Sign in to Foundry and open the Models experience
Marlene walks through accessing Microsoft Foundry from the lab browser, signing in, and navigating the UI to the build area. The focus is on reaching the model catalog to begin experimentation with Claude.
- •Open Foundry in browser and sign in using provided credentials
- •Navigate via ‘Start building’ → ‘Build’ toggle → ‘Models’
- •Access the model list and select Claude Sonnet 4.6
- 12:56 – 15:29
Test Claude in Foundry Playground with system prompts
The Foundry playground is used to verify the model is working and to demonstrate how system prompts change behavior. A playful cupcake persona prompt is used to illustrate rapid prompt iteration.
- •Playground enables quick chat-based testing and model comparison
- •Default assistant behavior verified with a basic ‘Hello’ exchange
- •System prompt updated to a themed persona (frosting puns)
- •Reinforces playground as the place to iterate on prompts before coding
- 15:29 – 17:29
Bring the Foundry model into VS Code: endpoint + API key configuration
The workshop transitions from UI testing to local development. Attendees copy the Target URI and API key, then update the project’s .env file with the correct endpoint format.
- •Copy Target URI and API keys from the model Details tab
- •Edit .env in VS Code to set endpoint, key, and model deployment
- •Critical fix: remove ‘/v1/messages’ so endpoint ends with ‘.../anthropic’
- •Prepared environment includes prebuilt files for quick start
- 17:29 – 20:01
Build a basic agent using Microsoft Agent Framework (Python)
Marlene introduces Microsoft Agent Framework as an open-source way to build agents and provides a code block to create a simple ‘cupcake agent.’ The agent is run locally to validate end-to-end connectivity to Claude.
- •Agent Framework introduced as Microsoft’s open-source agent SDK
- •Paste provided code into agent.py to define and configure the agent
- •Run via terminal command and test with a simple greeting
- •Validates .env configuration and model access from code
- 20:01 – 23:03
Add tools via MCP: connect to the cupcake-store MCP server
The session adds external capabilities using Model Context Protocol (MCP). By connecting an MCP server URL, the agent gains tool access to store data like available cupcake flavors.
- •MCP explained: standard for agents to talk to external systems
- •MCP server exposes tools, prompts, and resources
- •Add cupcake-store MCP server configuration in code
- •Attach MCP server as agent tools for real-time store information
- 23:03 – 24:05
Troubleshoot and verify tool access: querying available flavors
A brief debugging moment shows why saving code and properly wiring MCP matters: without it, the agent lacks store context. After saving and rerunning, the agent correctly lists available flavors from the MCP-backed toolset.
- •Initial ‘confusion’ response reveals MCP/tools weren’t active (file not saved)
- •Save changes, restart agent, and rerun the same query
- •Agent returns live flavor list (e.g., vanilla, lemon, red velvet, chocolate)
- •Demonstrates the value of tool-backed context vs. pure model guessing
- 24:05 – 26:40
Load reusable instructions and a welcome banner from MCP prompts
The agent is further refined by pulling a persona and greeting behavior from MCP-hosted prompts. This centralizes instruction management and standardizes the user experience without hardcoding prompts locally.
- •MCP prompts used as reusable instruction snippets
- •Code updated to fetch agent instructions + custom welcome banner
- •Restart agent to see updated greeting and guided flow
- •Highlights maintainability: prompts can be managed server-side
- 26:40 – 31:50
End-to-end ordering flow: customer registration, cupcake selection, voucher, pickup queue
Marlene demonstrates the completed agent experience: registering a new customer ID, selecting a cupcake, and entering a rotating voucher code displayed in the room. Orders appear on a dashboard and are approved for pickup by staff.
- •Agent collects user details to create a customer ID
- •User selects cupcake flavor by menu number
- •Voucher gate prevents overload; rotating code used to confirm orders
- •Dashboard shows ‘preparing’ and ‘ready for pickup’; staff approve orders
- 31:50 – 34:16
Wrap-up: where to get the workshop code, next steps, and Foundry learning resources
The session closes with pointers to the full workshop materials and ways to continue exploring Foundry. Marlene highlights free Azure credit, documentation, and Microsoft Learn courses for further practice.
- •Workshop/code available afterward for at-home practice
- •Speakers available for questions during/after the session
- •Next steps: search Microsoft Foundry, explore docs, use free Azure credit
- •Continue learning via Microsoft Learn courses and linked resources
