{"id":58353,"date":"2025-10-01T07:05:00","date_gmt":"2025-10-01T14:05:00","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/dotnet\/?p=58353"},"modified":"2025-11-04T10:02:05","modified_gmt":"2025-11-04T18:02:05","slug":"introducing-microsoft-agent-framework-preview","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/dotnet\/introducing-microsoft-agent-framework-preview\/","title":{"rendered":"Introducing Microsoft Agent Framework (Preview): Making AI Agents Simple for Every Developer"},"content":{"rendered":"

Building AI agents shouldn’t be rocket science. Yet many developers find themselves wrestling with complex orchestration logic, struggling to connect multiple AI models, or spending weeks building hosting infrastructure just to get a simple agent into production.<\/p>\n

What if building an AI agent could be as straightforward as creating a web API or console application?<\/p>\n

Agents and Workflows<\/h2>\n

Before we dive deeper, let\u2019s define the two core building blocks of agent systems: agents<\/strong> and workflows<\/strong>. We’ll start by looking at their core components and how they relate to one another<\/p>\n

Agents<\/h3>\n

Across the web, you\u2019ll find many definitions of agents. Some conflicting, some overlapping.<\/p>\n

For this post, we\u2019ll keep it simple:<\/p>\n

Agents are systems that combine reasoning, context, and tools to pursue objectives.<\/strong><\/p>\n

\"Diagram<\/p>\n

Let’s break down each of those components:<\/p>\n

    \n
  • Reasoning and decision-making<\/strong>: Agents need reasoning capabilities to decide which actions to take toward their objectives. Today, this is often powered by large language models (LLMs). However, other techniques such as search algorithms, planning systems, or classical AI approaches can also be used.<\/li>\n
  • Context awareness<\/strong>: Context is the external data or state that informs decision-making. Because models don\u2019t have built-in access to real-time or system-specific information, additional inputs like conversation history, knowledge bases, or enterprise data are required to make informed decisions.<\/li>\n
  • Tool usage<\/strong>: Tools are discrete, callable capabilities such as APIs, Model Context Protocol (MCP) tools, code execution, or data queries. They typically extend what a system can do but do not make decisions themselves. For example, a weather API provides data, but it does not decide how that data is used. Tools can be used to carry out actions or to gather additional context that informs future decisions.<\/li>\n<\/ul>\n

    Agents bring these components together to maintain awareness, make decisions, and act dynamically in pursuit of their goals.<\/p>\n

    Workflows<\/h3>\n

    As objectives grow in complexity, they need to be broken down into manageable steps. That\u2019s where workflows come in.<\/p>\n

    Workflows structure complex objectives into sequences of steps, coordinating tasks across people or systems to reach a goal efficiently.<\/strong><\/p>\n

    Imagine you’re launching a new feature on your business website. If it’s a simple update, you might go from idea to production in a few hours. But for more complex initiatives, the process might include:<\/p>\n

      \n
    • Requirement gathering<\/li>\n
    • Design and architecture<\/li>\n
    • Implementation<\/li>\n
    • Testing<\/li>\n
    • Deployment<\/li>\n<\/ul>\n

      \"Diagram<\/p>\n

      A few important observations:<\/p>\n

        \n
      • Each step may contain subtasks.<\/li>\n
      • Different specialists may own different phases.<\/li>\n
      • Progress isn\u2019t always linear. Bugs found during testing may send you back to implementation.<\/li>\n<\/ul>\n

        Success relies on planning, orchestration, and communication across stakeholders.<\/p>\n

        Agents + Workflows<\/h3>\n

        While workflows can function purely as predetermined sequences, integrating agents adds dynamic decision-making and adaptability, enabling more intelligent and autonomous process management.<\/p>\n

        Agents, workflows, and their underlying components are all highly composable:<\/p>\n

          \n
        • Agents can call multiple tools<\/li>\n
        • Tools may encapsulate agent-like behaviors in narrow scopes<\/li>\n
        • Workflows can sequence agents, tools, and other workflows<\/li>\n
        • Agents themselves may internally run workflows<\/li>\n<\/ul>\n

          Together, this layered architecture supports building flexible, intelligent, and scalable systems.<\/p>\n

          Meet Microsoft Agent Framework<\/h2>\n

          Microsoft Agent Framework is a comprehensive set of .NET libraries that reduces the complexity of agent development. Whether you’re building a simple chatbot or orchestrating multiple AI agents in complex workflows, Microsoft Agent Framework provides the tools you need to:<\/p>\n

            \n
          • Build<\/strong> agents with minimal boilerplate code<\/li>\n
          • Orchestrate<\/strong> multi-agent workflows with ease<\/li>\n
          • Host<\/strong> and deploy agents using familiar .NET patterns<\/li>\n
          • Monitor<\/strong> and observe agent behavior in production<\/li>\n<\/ul>\n

            For more details, see the Introducing Microsoft Agent Framework post in the Foundry blog<\/a>.<\/p>\n