Welcome to the world of AI orchestration with the Microsoft Agent Framework (MAF). This open-source, multi-language toolkit empowers developers to build, deploy, and manage intelligent agents with ease. One of its standout features is the durable workflow programming model, which lets you chain agents and tasks into robust, multi-step pipelines. Whether you're prototyping in a console app or scaling to Azure Functions, MAF handles execution, data flow, and error recovery — all while keeping your work resilient. In this article, we'll unpack ten key insights that every developer needs to know about durable workflows in MAF. From core building blocks to advanced patterns, you'll get a complete picture of how to leverage this powerful framework.
1. Workflows Are Directed Graphs of Executors
At the heart of every MAF workflow lies a directed graph of executors. Each executor is a self-contained unit of work that receives typed input, processes it, and emits output. You define these steps, then wire them together using the workflow builder API. The framework intelligently manages execution order, passes data between steps, and handles error propagation automatically. This graph-based approach supports sequential chains, parallel fan-out/fan-in patterns, conditional branching, and even human-in-the-loop approvals. The result is a flexible, maintainable pipeline that mirrors real-world business processes.
2. Getting Started Requires Only Two NuGet Packages
To begin building workflows, you need a .NET project and two NuGet packages: Microsoft.Agents.AI and Microsoft.Agents.AI.Workflows. Add them via the dotnet CLI or Package Manager. The core workflow runtime is lightweight, runs in-memory, and is perfect for local development and quick prototypes. Once installed, you can define executors, create a workflow builder, and run your first pipeline in minutes. No complex infrastructure — just code and the framework.
3. Executors Are the Fundamental Building Blocks
An executor is a class that inherits from Executor<TInput, TOutput> and overrides the HandleAsync method. Inside that method, you implement the actual logic — be it querying a database, calling an AI model, or sending an email. The executor is strongly typed, meaning input and output types are fixed at compile time, which reduces runtime errors. For example, an OrderLookup executor takes an order cancel request and returns an order object. This modular design makes testing and reuse straightforward.
4. The Workflow Builder Connects Executors into a Pipeline
After defining your executors, you use the workflow builder to link them together. The builder provides a fluent API for adding steps and defining dependencies. For instance, you can connect OrderLookup to OrderCancel and then to SendEmail, forming a linear chain. Alternatively, you can branch or fan out to multiple parallel agents. The builder ensures all data flows correctly and that the graph is valid before execution begins. This declarative approach separates orchestration logic from implementation.
5. The In-Process Runner Enables Local Testing
MAF includes a built-in in-process workflow runner that executes everything in memory. It’s ideal for development and unit testing because you don’t need an external service. You can run workflows directly in a console app, inspect intermediate results, and debug step by step. The runner supports all patterns — sequential, parallel, conditional — and simulates errors gracefully. Once your logic is solid, you can migrate to a durable backend (like Azure Storage) without changing executor code.
6. Durability Comes from Persisting Workflow State
Durable workflows survive process restarts, crashes, and long pauses by persisting their state to a durable store. MAF integrates with Azure Storage (tables and queues) to save the workflow’s progress. If an executor fails or the host reboots, the framework resumes exactly where it left off — no data loss. This is accomplished via event sourcing: each step’s outcome is recorded, and the workflow can be rehydrated from the stored events. Durability is key for production scenarios where reliability is non-negotiable.
7. Parallel AI Agents Can Be Orchestrated with Fan-Out/Fan-In
One powerful pattern is executing multiple agents in parallel using fan-out (broadcasting a message to several agents) and fan-in (aggregating their responses). For example, you might send a customer query to both a sentiment analyzer and a FAQ lookup agent, then merge results. MAF’s workflow builder lets you define such concurrent steps easily. The framework handles parallelism, waits for all branches, and then proceeds. This accelerates workflows that involve independent AI tasks.
8. Human-in-the-Loop Approvals Are Built-In
Workflows often require human judgment — like approving a cancellation or reviewing suspicious activity. MAF supports human-in-the-loop steps that pause execution until an external event (e.g., an API call from a user approval portal) arrives. The framework saves the workflow state durably while waiting, and resumes when the approval signal is received. This pattern integrates seamlessly with existing business processes and ensures compliance.
9. Hosting in Azure Functions Is Straightforward
To scale workflows in the cloud, you can host them inside Azure Functions. MAF provides bindings and triggers that let an Azure Function act as the workflow host. The durable storage (Azure Storage) decouples execution from the function instance, so workflows can run for hours or days. The function can kick off a workflow, deliver results, and even handle orchestrator failures. This serverless approach minimizes management overhead while maximizing resilience.
10. Error Propagation and Retry Policies Keep Workflows Robust
MAF workflows automatically propagate errors through the graph. If an executor throws an exception, the framework can reroute the workflow to a fallback path or halt execution. You can define custom retry policies — exponential backoff, maximum attempts — for transient failures. Together with durable state, these features ensure that temporary hiccups (like network timeouts) don’t permanently derail your pipeline. The result is a resilient system that meets the demands of enterprise AI applications.
Durable workflows in the Microsoft Agent Framework transform how you orchestrate AI agents. By breaking tasks into reusable executors, wiring them with a graph builder, and layering in persistence, you get reliability without sacrificing simplicity. Whether you're building a quick prototype or a production-grade solution, these ten insights give you a solid foundation. Experiment with the framework, try different patterns, and see how durable workflows can streamline your next AI project.