microsoft/agent-framework

What it solves

Microsoft Agent Framework (MAF) simplifies the transition from AI agent prototypes to production-grade systems. It addresses the complexity of orchestrating multi-agent workflows, ensuring durability, observability, and provider flexibility so that developers can evolve their architecture without needing to rewrite their entire codebase.

How it works

MAF provides a consistent API across Python and .NET (with a separate SDK for Go) to build and manage agents. It uses a graph-based orchestration system to handle complex patterns like sequential execution, concurrent tasks, handoffs between agents, and group collaboration. The framework includes a middleware system for request/response processing and supports multiple LLM providers, including Azure OpenAI and OpenAI.

Who it’s for

Developers and teams building production AI agents and multi-agent workflows who require more than simple stateless chat loops and need features like human-in-the-loop control, distributed tracing, and the ability to deploy to hosted infrastructure.

Highlights

• Multi-language support: Native implementations for Python and C#/.NET, with a Go SDK available.
• Graph-based workflows: Supports complex orchestration patterns including sequential, concurrent, and handoff patterns.
• Production-ready tooling: Built-in OpenTelemetry integration for observability, checkpointing for durability, and a DevUI for testing and debugging.
• Declarative agents: Ability to define agents using YAML for faster setup and versioning.
• Provider flexibility: Compatible with a broad ecosystem including Microsoft Foundry, Azure OpenAI, and OpenAI.