CHAPTERS
Session goal: build and deploy a Claude-powered agent in Microsoft Foundry
Marlene Mhangami opens the workshop and outlines the hands-on outcome: deploy a Claude model in Microsoft Foundry, wire it into an agent, and give it tools. The session is positioned as practical, with live building rather than theory-only discussion.
Why agentic systems change the requirements (reasoning, reliability, tools)
The talk frames the industry shift from single-turn chatbots to agentic systems that plan, reason, and act over time. This shift introduces new requirements around multi-step reasoning, enterprise reliability/observability/security, and integration with external tools and data.
What Microsoft Foundry provides: unified platform + connectors + enterprise controls
Marlene introduces Microsoft Foundry as a unified platform for building AI apps and agents at scale, integrating with common developer tools. She highlights core components (models, agent service, tools/integrations, ML services) and emphasizes enterprise-ready governance and security integrations.
Why use Claude in Foundry: reasoning quality + production pipeline
The session explains the benefits of pairing Claude models with Foundry. Claude provides strong planning/long-context reasoning, while Foundry adds agent orchestration, tool access, and a faster path from prototype to production with monitoring and evaluation tools.
Workshop kickoff: launch the guided lab environment (Skillable)
Participants are directed to the workshop URL and guided to launch a prepared environment. The lab is designed for self-paced progress with step-by-step instructions, plus helper features that auto-fill fields to reduce typing.
Scenario setup: Sparkles cupcake shop needs an ordering agent
The workshop’s narrative sets up a cupcake store overwhelmed by customers and flavors. The goal is to build an agent that can manage customer intake and cupcake ordering, using tools to access store data and process orders.
Get into Foundry: sign in, pick project, and open the Models view
Marlene walks through signing into Foundry inside the lab browser, then navigating to the Build area and selecting Models. This positions participants to choose a Claude model and begin experimenting.
Use the Foundry playground: test Claude + system prompts
In the model playground, participants chat with Claude Sonnet and see how system instructions shape behavior. A playful “sentient cupcake” system prompt demonstrates prompt control and quick iteration in the UI.
Bring the model to VS Code: endpoint/URI and API key configuration
The workshop transitions from UI experimentation to local development. Participants copy the target URI and API key from Foundry, then update the project’s .env file—highlighting a critical fix: removing the trailing “v1/messages” so the endpoint ends correctly.
Build a basic agent with Microsoft Agent Framework (Python)
Participants implement a minimal “Cupcake agent” using Microsoft Agent Framework in Python. The agent uses the configured environment variables to connect to Claude and respond in a simple terminal chat loop.
Add tools via MCP: connect to the cupcake-store MCP server
The agent is upgraded to use Model Context Protocol (MCP) to access external tools and store data. Participants connect to a cupcake-store MCP server URL, pass it as a tool source to the agent, and learn by debugging a common mistake (forgetting to save the file).
Load persona + UX from MCP prompts: instructions and welcome banner
Participants pull reusable prompts from the MCP server to standardize the agent’s behavior and greeting. The agent now displays a custom banner and follows a guided intake flow to create a customer profile before ordering.
End-to-end cupcake ordering flow: ID creation, flavor selection, voucher, dashboard
The live demo walks through registering a new customer, selecting a cupcake flavor, and validating via a rotating voucher code shown on a dashboard. Orders then appear in a “preparing” and “ready for pickup” queue, with staff approving orders at the front.
Wrap-up: resources, code availability, and how to continue with Foundry
Marlene closes by pointing attendees to documentation, Microsoft Learn courses, and free credits for continued experimentation. She notes the workshop/code will be available later and invites questions and follow-up conversations during the event.
