Microsoft AI Agentic Workshop: Hands-On Multi-Agent Patterns for Azure

What the Workshop Teaches and Who Should Use It

The repository is workshop material, not a standalone application. Its audience is developers, architects, and teams evaluating how to build production-ready agentic AI systems on Azure. The core deliverable is a working end-to-end architecture: a backend database, an MCP tools server, agent orchestration using the Microsoft Agent Framework, an application backend, and a React or Streamlit frontend. The workshop presents a concrete business scenario (documented in SCENARIO.md) that participants use as the context for all exercises. The focus is on comparing different multi-agent patterns within the same business scenario rather than implementing a generic chatbot. Three orchestration patterns are available: single-agent, Magentic multi-agent orchestration, and handoff-based domain routing. Teams that want framework-agnostic patterns or that do not have Azure subscriptions will find the workshop's Azure dependencies restrictive.

Three Orchestration Patterns in the Microsoft Agent Framework

The Microsoft Agent Framework integration in the workshop covers three patterns, documented in agentic_ai/agents/agent_framework/README.md. The single-agent pattern runs one agent against the full set of MCP tools. The Magentic multi-agent orchestration pattern coordinates multiple specialized agents using a fan-out topology where a planner agent breaks a request into subtasks and assigns each to a domain-specific agent. The handoff-based domain routing pattern routes requests between agents based on which domain the request belongs to, with explicit handoff events tracked in the session state. Each pattern has a corresponding implementation in the repository. The workshop is structured so participants can implement all three and compare how the choice of pattern affects latency, cost, and error handling for the same business scenario.

Durable Fraud Detection Workflow and Human-in-the-Loop

The workflow directory contains a fraud detection demo that illustrates hybrid Workflow plus Durable Task architecture. This demo at agentic_ai/workflow/fraud_detection_durable/ runs a fan-out/fan-in topology: multiple parallel analysis tasks execute simultaneously, their results merge at a synchronization point, and then a human-in-the-loop approval step pauses the workflow until a human confirms or rejects the fraud decision. The Durable Task pattern means the workflow persists through process restarts: if the service crashes between the parallel analysis phase and the approval step, the workflow resumes from the last checkpoint rather than starting over. The workshop uses this demo to teach checkpoint-based resilience, which is a practical requirement for long-running agentic workflows that interact with external systems and need audit trails.

Observability and the Grafana Dashboard Integration

The workshop includes an observability module in agentic_ai/observability/ that connects all agent executions, tool calls, and LLM invocations to Azure Application Insights. The setup guide at agentic_ai/observability/README.md documents how to configure the tracing integration. Pre-built Grafana dashboards are provided for visualizing the telemetry. This means workshop participants get a working observability stack without building one from scratch. For production agentic systems, this matters because an agent failure is often not visible from the outside: the request succeeds but the agent took the wrong path, called the wrong tool, or incurred unexpected cost. Application Insights traces each step and the Grafana dashboards surface patterns across runs.

Deploying to Azure: Developer CLI, Terraform, and Enterprise Security

The workshop offers three deployment paths. The Azure Developer CLI (azd) path is a single-command quick start that provisions all resources automatically. The manual path uses PowerShell with Terraform or Bicep templates. The enterprise security profile adds VNet integration, private endpoints, and managed identity to the base deployment. A GitHub Actions CI/CD pipeline at .github/workflows/ automates dev-to-production promotion with per-developer environments, OIDC authentication, agent evaluation gates, and filtering for documentation-only changes. The infrastructure code is in the infra/ directory. This multi-path deployment design means a developer can start quickly with the azd path and then move to the Terraform path when they need more control over the resource configuration.

Limitations and What the Workshop Does Not Cover

The workshop is built entirely on Azure and the Microsoft Agent Framework. Developers who need patterns for AWS Bedrock, Google Vertex AI, or LangChain will not find directly applicable infrastructure code. The Agent Framework integration depends on the Microsoft Agent Framework repository at github.com/microsoft/agent-framework, which is a separate project. If that project changes its API, the workshop code may break. The repository has no GitHub releases, so there is no versioned snapshot of a known-working configuration. The workshop is under active development as of the last push on 2026-09-24, which means the code is current but the patterns may change as the Microsoft Agent Framework evolves.

Compared to LangChain Tutorials and OpenAI Cookbook, and Licence Terms

The closest alternatives are the LangChain documentation examples and the OpenAI Cookbook. Both provide agent patterns and multi-agent examples, but neither includes production-grade Azure infrastructure code, Terraform templates, or a CI/CD pipeline with evaluation gates. The OpenAI Cookbook is framework-agnostic and covers a broader range of providers. The Microsoft AI Agentic Workshop is narrower in scope but deeper on the Azure deployment and enterprise security side. A team that primarily uses OpenAI's API directly without Azure infrastructure is better served by the OpenAI Cookbook. The workshop is MIT-licensed, which permits use and modification without restrictions on commercial adaptation. The repository includes a CODE_OF_CONDUCT.md and SECURITY.md, consistent with Microsoft open-source project governance.

Editorial conclusion

The Microsoft AI Agentic Workshop is the right starting point for developers and solution architects building production agentic AI on Azure who want reference patterns with real infrastructure code rather than documentation-only guides. The workshop covers the full stack from database to frontend and includes a CI/CD pipeline, which is rare in workshop repositories. The dependency on Azure and the Microsoft Agent Framework makes it less suitable for teams on other cloud providers or using frameworks like LangChain. Verify Azure subscription costs before deploying the enterprise security profile, as the VNet and private endpoints add significant infrastructure expense. The repository is MIT-licensed. The last push was on 2026-09-24.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. microsoft/OpenAIWorkshop on GitHub
4. README