Microsoft Agent Framework

Microsoft Agent Framework (MAF) is an open-source, multi-language framework from Microsoft for building AI agents and multi-agent workflows in Python and .NET. The README describes it as aimed at teams moving agents from prototype to production, with samples and hosting patterns for local development and cloud deployment. A separate Go SDK is maintained in its own repository.

What It Is

MAF is an agent development SDK. Developers create agents backed by an LLM provider, give them instructions, tools, and memory, and compose them into graph-based workflows. It is installed as the agent-framework Python package or the Microsoft.Agents.AI NuGet package. The quickstart creates an agent against Microsoft Foundry and runs it with a prompt.

Orchestration and Workflows

Workflows support sequential, concurrent, handoff, and group collaboration patterns. The documentation lists checkpointing and resuming, streaming, human-in-the-loop control, and time-travel. Agents can be used as steps within workflows. A middleware system handles request/response processing, exception handling, and custom pipelines.

Agent Capabilities and Tooling

Documented capabilities include function tools, structured output, multi-turn conversations, persisted conversations and memory, RAG, agent skills, background agents, and an agent harness. Agents can also be defined declaratively in YAML. Built-in OpenTelemetry integration supports tracing and debugging, and DevUI provides an interactive interface for developing and testing agents and workflows. AF Labs holds experimental packages such as benchmarking and reinforcement learning.

Providers and Hosting

The framework supports multiple LLM providers, including Microsoft Foundry, Azure OpenAI, OpenAI, and the GitHub Copilot SDK. Hosting options include Foundry hosted agents, A2A, self-hosting, and Azure Functions with durable hosting. Migration guides cover moving from AutoGen and Semantic Kernel.