Building Enterprise Agents with Microsoft Foundry and Azure PaaS

3:00 오후 - 4:00 오후 (UTC)

AI agents are rapidly moving beyond simple conversational experiences. In this session, we will build the foundation of a production-ready agent using Microsoft Foundry Agent Service, Microsoft Agent Framework, and Azure application services. Starting from a practical enterprise operations scenario, we will define the agent’s responsibilities, instructions, tools, conversation state, and structured outputs. We will then connect the agent to a Foundry model, add a typed function tool, stream responses to a web application, and examine the architectural differences between a model call, an assistant, an agent, and an agent workflow. The session will also explore hosting options, including Foundry hosted agents and custom agent applications deployed to Azure Container Apps. Along the way, we will discuss identity, configuration, state management, failure boundaries, and the design decisions that separate a useful prototype from an application that can evolve toward production. By the end of the session, attendees will understand the core components of a Microsoft Foundry agent application and will have a reusable starting point for building their own agentic solutions. Key topics: Microsoft Foundry Agent Service Microsoft Agent Framework Agent instructions and tool contracts Conversation state and streaming responses Structured outputs Hosted agents and Azure Container Apps Production-oriented agent architecture

3:00 오후 - 4:00 오후 (UTC)

Adding more agents does not automatically create a better system. Multi-agent solutions need clear responsibilities, controlled execution, recoverable state, and carefully designed decision boundaries. In this session, we will extend the agent built in Episode 1 into a coordinated incident-management workflow. Specialist agents will handle triage, knowledge retrieval, impact assessment, and remediation planning. We will use Microsoft Agent Framework to demonstrate sequential, concurrent, and handoff orchestration patterns, and discuss where each pattern is appropriate. The workflow will introduce typed state, checkpointing, failure recovery, and human-in-the-loop approval before any sensitive remediation action can continue. We will also separate deterministic application logic from agentic reasoning, showing which decisions should be handled by an AI agent and which should remain enforced by conventional code. Attendees will see how a multi-agent workflow behaves when interrupted, how it resumes from saved state, and how human approval can be incorporated without losing execution context. By the end of the session, attendees will be able to design controlled multi-agent workflows that combine AI judgment with reliable business-process execution. Key topics: Multi-agent orchestration patterns Sequential, concurrent, and handoff workflows Typed workflow state Checkpointing and recovery Human-in-the-loop approvals Agentic versus deterministic steps Failure handling and idempotency. This session is a part of a series, learn more here! https://aka.ms/ProBadgeSeries-TY [eventID:27508]

3:00 오후 - 4:00 오후 (UTC)

An enterprise agent becomes genuinely useful when it can act on business systems. It also becomes significantly more dangerous if those integrations are not governed correctly. In this session, we will connect our Microsoft Foundry agent to real enterprise capabilities using Azure API Management, Azure Logic Apps, Azure Functions, Azure Service Bus, OpenAPI tools, and Model Context Protocol servers. The agent will create an incident ticket, initiate an approval workflow, invoke a protected business API, submit a long-running remediation command, and report the final result back to the user. Rather than allowing the model direct access to backend systems, we will introduce a governed integration layer with typed tool contracts, authentication, authorization, rate limiting, schema validation, retries, and centralized logging. We will also compare the roles of OpenAPI tools, MCP servers, Logic Apps actions, Azure Functions, and asynchronous messaging. Managed identities and Microsoft Entra ID will be used to eliminate embedded API keys and apply least-privilege access across the solution. By the end of the session, attendees will understand how to give agents meaningful enterprise capabilities without bypassing security, integration governance, or operational controls. Key topics: Microsoft Foundry tools OpenAPI and Model Context Protocol Azure API Management as an AI gateway Azure Logic Apps agent actions Azure Functions Azure Service Bus Managed identity and Microsoft Entra ID Secure and asynchronous tool execution. This session is a part of a series, learn more here! https://aka.ms/ProBadgeSeries-TY [eventID:27509]

3:00 오후 - 4:00 오후 (UTC)

A successful demonstration does not prove that an agent is ready for production. Production readiness requires measurable quality, end-to-end observability, repeatable testing, controlled releases, and a clear operational model. In the final session, we will evaluate and operate the agent solution built throughout the series. We will create a representative evaluation dataset covering normal requests, ambiguous cases, tool failures, unauthorized actions, outdated knowledge, and prompt-injection attempts. Using Microsoft Foundry evaluation and observability capabilities, we will measure task completion, tool-call accuracy, groundedness, relevance, latency, safety, and operational reliability. We will trace requests across the agent, API Management, Logic Apps, Azure Functions, Service Bus, and the worker application using distributed tracing and Application Insights. During the live demonstration, we will deliberately introduce a faulty tool definition, observe the resulting regression, identify the cause through traces and evaluation results, correct the implementation, and apply a quality gate before releasing the updated version. The session will conclude with production monitoring, alerting, versioning, rollout and rollback strategies, and agent governance across multiple teams and environments. By the end of the session, attendees will know how to move from “the agent appears to work” to evidence that it is reliable, explainable, supportable, and safe to operate. Key topics: Microsoft Foundry evaluation Agent and tool-call quality metrics Distributed tracing Application Insights Regression testing safety and adversarial testing CI/CD quality gates Monitoring, incident response, and governance. This session is a part of a series, learn more here! https://aka.ms/ProBadgeSeries-TY [eventID:27510]