Microsoft Agent Framework

Microsoft Agent Framework is the open-source, multi-language framework for building production agentic AI applications in Python and C#/.NET. It exposes consistent agent abstractions, enterprise-grade state management, type safety, middleware, OpenTelemetry-based observability, and orchestration patterns covering sequential, concurrent, handoff, and group-collaboration topologies. The framework supports multiple model providers — Microsoft Foundry, Azure OpenAI, and OpenAI — and supports both local-development and cloud-hosted execution on Foundry infrastructure.

The framework is the direct successor to Semantic Kernel and AutoGen, consolidating lessons from both into a single platform aimed at the gap between prototype agents and production systems. Durable agent execution with checkpointing and restartability, declarative agent definitions in YAML, and first-class human-in-the-loop control let teams build, test, and operate multi-agent systems with the same engineering rigor they expect from other server-side software. It is the recommended foundation for any team building agentic AI on Microsoft’s stack, and it underlies many of the agent systems Microsoft itself ships.