Microsoft has launched a new platform that combines two of its major AI tools, aiming to make building AI applications easier. This new system, called the Microsoft Agent Framework, merges the capabilities of Semantic Kernel and AutoGen. Both tools are open source and available on GitHub, allowing developers to experiment and build their own AI workflows.

The goal is to create a flexible environment where different AI models and services can work together seamlessly. Microsoft has quickly added support for new practices like Model Context Protocol (MCP) and Agent2Agent communication. Users can pick the AI models they want, whether from Microsoft or other providers.

Building AI Workflows with the New Framework

The core idea behind the Microsoft Agent Framework is workflow orchestration. This means designing sequences or groups of AI tasks that work together to complete complex jobs. The framework supports two main types of orchestration. One is called workflow orchestration, which builds on Semantic Kernel. It helps automate business processes by calling a chain of agents, each performing a specific task. Prompts are constructed with predefined formats, and results from earlier steps are used to inform later ones.

The other type is agent orchestration, inspired by AutoGen. This approach allows for more dynamic workflows where agents can interact in real-time based on open-ended prompts. Both types can be combined or embedded within each other.

Innovative Orchestration Models for AI Agents

One of the most exciting features is the variety of orchestration models. The simplest is sequential orchestration—agents are called one after the other, waiting for each response before moving on. This is straightforward but effective for many tasks. More advanced is concurrent orchestration, where multiple agents are called simultaneously, speeding up the process.

Microsoft has introduced new models that use language models in more sophisticated ways. Group chat orchestration allows agents to talk among themselves, sharing results until they reach a consensus. Hand-off orchestration takes sequential steps further by updating prompts based on previous responses, making workflows more adaptable.

The most innovative is what Microsoft calls “magentic” workflows. These involve a supervisory agent that manages a subset of other agents. This supervisor can coordinate tasks and bring in human input if needed. This setup is designed for complex problems that require more oversight and multiple decision points.

Building and Migrating to the New Platform

Creating new applications with the Agent Framework is designed to be straightforward, especially for developers familiar with .NET. You’ll need .NET 9 or later and access to AI models, either locally or in the cloud. Microsoft has made it easy to get started by offering packages like Microsoft.Agents.AI, which include tools for connecting to models through a simple API.

Developers can build agents starting with a chat interface that connects to their chosen model. They can set up prompts and run agents asynchronously, making the whole process flexible and scalable. The framework’s abstraction layer allows swapping models easily, whether they’re cloud-based or local, without changing the core code.

Existing code from Semantic Kernel and AutoGen can also be migrated, though it requires some adjustments. For example, moving from plug-ins to tools and working with external APIs means rewriting parts of the code. Microsoft emphasizes that this migration, while not instant, is a credible pathway for users wanting to upgrade their systems.

What the Future Holds for AI Workflows

Microsoft’s goal is for the Agent Framework to serve as a bridge across all its AI products. Whether you’re using Copilot Studio, Azure AI Foundry, or Microsoft 365 tools, the new platform aims to unify how AI agents are built and managed. The support for OpenAPI makes it easy to integrate external services, further expanding possibilities.

The framework also supports running agents anywhere—on-premises, in the cloud, or in containers—which offers great flexibility for enterprise deployments. Microsoft’s focus on local models, like those that run on NPUs, could open new doors for privacy and speed, especially as more small language models become available.

In the end, Microsoft is pushing to make AI workflows more dynamic, adaptable, and integrated. Building and managing these workflows will become more accessible, even as they grow in complexity. As AI continues to evolve, tools like the Microsoft Agent Framework are likely to play a big role in shaping how businesses automate tasks and solve complex problems.

Inspired by

• https://www.infoworld.com/article/4069808/unpacking-the-microsoft-agent-framework.html

Sources

• gmpg.org
• microsoft.com
• github.com
• github.com
• microsoft.com