A .NET Agent Workflow Orchestration Framework Based on Microsoft Agent Framework

This article introduces the open-source project agent-workflow-orchestrator-dotnet-maf, which is built on Microsoft Agent Framework (MAF) and aims to provide production-grade Agentic workflow orchestration capabilities for the .NET ecosystem, supporting the deployment of complex agent collaboration processes on Azure. The project is maintained by Gabegi, with source code hosted on GitHub (link: https://github.com/Gabegi/agent-workflow-orchestrator-dotnet-maf), and the update date is 2026-05-28.

With the improvement of LLM capabilities, AI applications are evolving towards autonomous agents, but deploying Agentic workflows faces challenges such as complex state management, difficult tool coordination, lack of multi-agent collaboration, and integration with enterprise tech stacks. As an official framework, Microsoft Agent Framework (MAF) provides standardized agent building capabilities for the .NET ecosystem, abstracting core functions like lifecycle management and tool calls to help developers focus on business logic.

agent-workflow-orchestrator-dotnet-maf is an open-source workflow orchestration system based on MAF, supporting process control structures such as sequential execution, conditional branching, and parallel processing. The core design goal is to separate workflow definition from the execution engine: developers define the structure through declarative configuration or code, while the orchestrator handles state persistence, error recovery, and concurrency control, ensuring consistent behavior across environments from local to Azure Kubernetes clusters.

1. MAF Agent Lifecycle Management: Each workflow node corresponds to a MAF agent instance. The framework manages its creation/config/destruction. Agents communicate via message interfaces (synchronous/asynchronous), and dependency injection integration facilitates access to services like LLM clients and vector databases.
2. State Persistence & Recovery: Uses Azure Blob Storage or Cosmos DB to save execution state. An incremental snapshot strategy balances recovery capability and performance, supporting resumption from breakpoints.
3. Multi-agent Collaboration: Supports Master-Worker and Peer-to-Peer modes. Dynamic task allocation based on capabilities (agents declare task types and capability scores during registration) enables elastic scaling.

1. Enterprise Process Automation: Transform traditional processes (e.g., insurance claims), where each step is handled by specialized agents. Adapts to changes in business rules and processes unstructured inputs (emails/call records).
2. Multi-step Code Generation & Review: Build intelligent programming assistants covering requirement analysis, code generation, test writing, and review. Closed-loop feedback improves code quality.
3. Research & Data Pipelines: Orchestrate complex analysis workflows (data acquisition, cleaning, model running, report generation), handling dependencies and failure retries.

For deployment, Azure Container Apps or AKS are used to host the orchestrator, leveraging auto-scaling to handle load fluctuations. Observability is achieved via Azure Monitor and Application Insights for full-link tracing. For security, it integrates Azure Managed Identity for authentication, stores sensitive configurations in Azure Key Vault, and uses log desensitization to protect privacy.

This project combines MAF's standardized capabilities with Azure cloud-native services to provide a scalable, observable, and recoverable agent orchestration platform for the .NET ecosystem. Future directions include: auto-optimizing workflow structure based on execution history, supporting more complex nesting/recursive patterns, deep integration with more enterprise systems, and providing a reference paradigm for AI-native application architectures.