microsoft/OpenAIWorkshop

What is this?

A free, open-source collection of code, docs, and sample scenarios for learning how to design, build, and deploy AI agents (software that can take actions on its own, not just chat) using Microsoft's AI tools. It's built as a self-guided workshop: you follow the business scenario, try the sample code, and experiment with different agent designs.

What can you actually do with it?

• Design and prototype AI agents for realistic business problems
• Compare different agent setups — a single agent, several agents working together, or agents that hand off tasks to each other
• Build a complete system — from a database and backend tools, through agent orchestration, to a web frontend (React or Streamlit)
• Deploy to Azure with enterprise-grade security (private networks, managed identities, automated CI/CD)

Main technologies inside

• Microsoft Agent Framework — handles orchestration of multiple agents, handoffs between them, and checkpointing
• MCP (Model Context Protocol) Server — a standard way to give agents extra tools (like database access) with authentication and role-based access control
• Workflow orchestration — a durable task system with "fan-out/fan-in" patterns (parallel tasks that merge), human approval steps, and real-time monitoring; includes a fraud-detection demo
• Observability — full tracing of what agents do, which tools they call, and how AI models are invoked, with pre-built Grafana dashboards
• Agent evaluations — test agent performance with custom metrics and datasets
• State persistence — store conversation history and agent state in memory or in CosmosDB
• Infrastructure-as-code — Terraform/Bicep templates for production deployment
• CI/CD pipelines — automated promotion from dev to production, with per-developer environments and evaluation gates

How to get started

The README gives a 6-step path:

1. Review setup instructions
2. Read the business scenario and agent design
3. Choose a multi-agent pattern (single, Magentic orchestration, or handoff)
4. Try the durable fraud-detection workflow with human-in-the-loop
5. Study the system architecture before customizing
6. Use the support guide for troubleshooting

Deployment options

Four ways to deploy, all documented:

• Azure Developer CLI — one-command quick start
• Manual deployment — PowerShell with Terraform/Bicep
• Enterprise security — VNet, private endpoints, managed identity
• GitHub Actions CI/CD — with OIDC authentication

Notes

• Licensed under the terms in the LICENSE file
• Has a Code of Conduct and Security Guidelines
• Microsoft explicitly links to the Agent Framework project for further reference