Microsoft has taken a bold step by merging two distinct AI development platforms—Semantic Kernel and AutoGen—into what is now called the Microsoft Agent Framework. Semantic Kernel brought a production-ready AI workflow engine to the table, while AutoGen contributed research-driven multi-agent capabilities that allowed developers to build complex agent systems without writing extensive code. The merger aims to give developers a unified platform that leverages the strengths of both approaches.
Both Semantic Kernel and AutoGen were open source, and Microsoft has continued that tradition with the new framework. Hosted on GitHub, the Microsoft Agent Framework includes sample projects and a ready-to-try Codespace virtual environment. This open approach encourages experimentation, lets developers adopt emerging best practices quickly, and provides flexibility in selecting AI models and providers. Features like Model Context Protocol (MCP) and Agent2Agent support are already built in, helping developers integrate the framework into diverse AI workflows.
At the core of the framework is workflow orchestration. Building on Semantic Kernel, this approach allows developers to define structured business processes and logic, chaining together multiple agents in a predictable manner. Prompts are constructed using predefined formats, and outputs from one agent can be automatically fed into the next, ensuring smooth end-to-end workflows. This makes it ideal for automating established business tasks where consistency and reliability are critical.
Complementing workflow orchestration is agent orchestration, which draws on AutoGen’s dynamic capabilities. This approach enables the creation of agent chains in real time, driven by open-ended prompts. Agents can spawn other agents and collaborate autonomously to solve tasks, making it suitable for exploratory workflows and adaptive problem-solving. By embedding workflow and agent orchestration together, the Microsoft Agent Framework provides developers with a versatile toolkit that supports both structured and dynamic AI applications.