Skip to main content

Build, observe, and govern multi-agent AI on Azure App Service

Multi-agent AI apps are easy to prototype and hard to run in production. Once several agents call language models and external tools on a user's behalf, you need to answer two operational questions: what are the agents doing? and what are they allowed to do? This learning path answers both.

You carry one app - a six-agent travel planner built with the Microsoft Agent Framework (MAF)

  • from a first deploy all the way to a fully observable, policy-governed app on Azure App Service. You never rewrite the app. In each step you add one operational capability around it: first deployment, then observability, then governance.
One app, carried across steps

Each step builds on the last and uses the same resources. Do the steps in order and keep your resources running until you reach Clean up at the end of the path. This path is based on the three-part blog series Multi-Agent AI on Azure App Service.

Meet the app​

The sample is a travel planner that turns a single request - a destination, dates, budget, and interests - into a complete day-by-day itinerary. It does this with six specialized agents that collaborate:

  • Travel Planning Coordinator - orchestrates the workflow and synthesizes the final plan.
  • Currency Conversion Specialist - converts budgets using a live exchange-rate tool.
  • Weather & Packing Advisor - checks forecasts and alerts using a weather tool.
  • Local Expert & Cultural Guide - adds local knowledge and etiquette tips.
  • Itinerary Planning Expert - builds the day-by-day schedule.
  • Budget Optimization Specialist - keeps the plan within budget.

The app runs on App Service as an ASP.NET Core API plus a continuous WebJob. The API accepts a request and returns immediately with a task ID; the WebJob runs the multi-agent workflow in the background and writes the result to Cosmos DB, which the client polls. Azure OpenAI provides the model; Service Bus decouples the API from the workflow.

What you will build​

By the last step, the travel planner is not just running - it is observable and governed:

  • Step 1 - Deploy the app to App Service and watch six agents produce a travel plan.
  • Step 2 - Observe every agent, token, and tool call in the Application Insights Agents (Preview) view, powered by OpenTelemetry GenAI semantic conventions.
  • Step 3 - Govern the agents by adding the Agent Governance Toolkit so a policy file decides which tools each agent may call - and denied calls are blocked at runtime.

What you need​

The app for this path lives in Azure-Samples/app-service-multi-agent-maf-otel. Step 1 walks you through cloning and deploying it. You start from the start branch (before governance) and add governance yourself in Step 3; the main branch holds the finished version if you want to compare.

The path​

Work through the steps in order. Use the checkboxes to track your progress - they are saved in your browser.

0 of 3 available steps complete
  1. Deploy the multi-agent app
    Scenarios & modern apps30-40 min

    Deploy the six-agent travel planner to App Service with the Azure Developer CLI.

  2. Observe the agents
    Monitoring & diagnostics20-30 min

    Trace every agent, token, and tool call in the Application Insights Agents view.

  3. Govern the agents
    Security & identity30-40 min

    Enforce policies on agent tool calls with the Agent Governance Toolkit.

Clean up​

When you finish the path (or want to stop), delete the single resource group azd created to remove every resource and stop billing. The final step of the path includes the exact command. Because azd up puts everything in one resource group, one delete removes it all.