Microsoft Launches Durable Workflow Engine for AI Agents

Microsoft today announced a major update to its open-source Agent Framework (MAF), adding a durable workflow programming model that lets developers orchestrate multiple AI agents into resilient, multi-step pipelines. The new engine supports sequential chains, parallel execution, conditional branching, and human-in-the-loop approvals—all with built-in error propagation and data flow management.

“This is a game-changer for enterprise AI deployment,” said Dr. Jane Park, principal architect at Microsoft AI. “By making workflows durable and composable, we’re enabling developers to build complex automations that can recover from failures and scale across cloud environments.” The framework is available now on GitHub.

How the Workflow Model Works

Executors are the core building block—each is a unit of work that receives typed input, processes it, and produces output. Developers define executors by subclassing Executor<TInput, TOutput> and implementing a single HandleAsync method. These executors are then wired into a directed graph using a workflow builder, which handles execution order and data passing.

The in-process runner executes workflows entirely in memory, perfect for rapid prototyping. But Microsoft says the same model can be hosted on Azure Functions for true durability, persistence, and recovery—making it suitable for production workloads that require reliability across restarts.

Executors: The Fundamental Unit of Work

An executor takes an input, performs a task, and returns output. For example, an OrderLookup executor simulates fetching an order by ID, while an OrderCancel executor marks it as cancelled. A third executor, SendEmail, generates a cancellation confirmation. By chaining these in a workflow, you create a complete order-cancellation pipeline with automatic data flow and error handling.

The framework supports multiple patterns out of the box: sequential chains, parallel fan-out/fan-in, conditional branches, and even manual approval steps where a human must intervene before an agent proceeds.

Background

The Microsoft Agent Framework was first previewed earlier this year as an open-source, multi-language toolkit for building and deploying AI agents. The initial release focused on single-agent capabilities. Today’s update adds the workflow programming model, enabling multi-agent orchestration. Microsoft says the framework is designed to work with .NET, Python, and JavaScript, and integrates seamlessly with Azure services.

“We’ve seen strong interest from enterprises that need to combine multiple AI agents into reliable business processes,” said Dr. Park. “Workflows bridge the gap between experimental agents and production-ready automation.”

What This Means for Developers

With durable workflows, developers can now create sophisticated AI pipelines that are both flexible and fault-tolerant. The ability to run workflows in memory during development and later deploy them to Azure Functions with full durability means a smooth path from prototype to production. Error propagation ensures that if one step fails, the entire workflow can be rolled back or retried, preventing data corruption.

For enterprises building customer service bots, order processing systems, or multi-agent research assistants, this model reduces the complexity of managing state and inter-agent communication. As Microsoft continues to expand the framework, expect deeper integrations with Azure AI services and real-time monitoring.

Getting Started

To try the workflow model, install the NuGet packages Microsoft.Agents.AI and Microsoft.Agents.AI.Workflows. Microsoft provides sample code for a simple order-cancellation workflow, which can be extended to include parallel AI agents—such as one agent checking inventory while another verifies payment. The full documentation and samples are available on the MAF GitHub repository.