On the coronary heart of Microsoft’s AI software improvement technique is Semantic Kernel, an open source set of instruments for managing and orchestrating AI prompts. Since its launch as a method to simplify constructing retrieval-augmented generation (RAG) purposes, it has grown right into a framework for constructing and managing agentic AI.

At Ignite in 2024, Microsoft introduced a number of new options for Semantic Kernel, positioning it as its most popular device for constructing large-scale agentic AI purposes. That announcement fashioned the premise of Semantic Kernel’s 2025 road map, with the primary parts already being delivered.

One of many extra necessary new options in Semantic Kernel is the Agent Framework, which can quickly transfer out of preview into common availability. It will guarantee a secure, supported set of instruments able to ship production-grade enterprise AI purposes. The Agent Framework will kind the premise of Semantic Kernel’s deliberate integration with Microsoft Research’s AutoGen, together with the discharge of a standard runtime for brokers that’s constructed utilizing each platforms.

Constructing agentic workflows with the Agent Framework

The Agent Framework is meant to assist build applications around agent-like patterns, providing a manner so as to add autonomy to purposes and to ship what Microsoft calls “goal-oriented purposes.” This can be a good definition of what trendy agentic AI ought to be: a manner of utilizing AI instruments to assemble and handle a workflow primarily based on a consumer request. It then permits a number of brokers to collaborate, sharing information and managing what may be regarded as lengthy transactions that work throughout many alternative software APIs and endpoints.

Obtainable as an extension to the bottom Semantic Kernel, the Agent Framework is delivered as a set of .NET libraries, which assist handle human/agent interactions and supply entry to OpenAI’s Assistant API. It’s supposed to be managed through dialog, although it’s straightforward sufficient to construct and run brokers that reply to system occasions moderately than direct human actions (and so as to add human approval steps as a part of a dynamic workflow). This allows you to concentrate on utilizing brokers to handle duties.

Semantic Kernel’s agent options are designed to increase the ideas and instruments used to construct RAG-powered AI workflows. As at all times, Semantic Kernel is how each the general orchestration and particular person brokers run, managing context and state in addition to dealing with calls to AI endpoints through Azure AI Foundry and comparable companies.

Brokers speaking to brokers

Building a Semantic Kernel agent requires an Agent class earlier than utilizing an Agent Chat to help interactions between your agent workflow and the AI and API endpoints used to finish the present activity. If a number of brokers must be known as, you need to use an Agent Group Chat to handle these inner prompts by utilizing Semantic Kernel to work together and move outcomes between one another. An Agent Group Chat may be dynamic, including and eradicating participant brokers as wanted.

You’re capable of construct on present Semantic Kernel methods, too. For instance, brokers can use present or new plug-ins in addition to name features. Working with exterior purposes is vital to constructing enterprise brokers, as they want to have the ability to dynamically generate workflows round each people and software program.

Having Semantic Kernel handle brokers ensures you possibly can handle each directions and prompts for the large language model (LLM) you’re utilizing, in addition to management entry to the APIs. Your code can handle authorization as crucial and add plug-in objects. Your plug-ins will handle API calls, with the agent setting up queries by parsing consumer inputs.

No-code brokers with AutoGen

Semantic Kernel’s integration with AutoGen builds on its Process Framework. That is designed to handle long-running enterprise processes and works with distributed software frameworks equivalent to Dapr and Orleans. Workflows are event-driven, with steps constructed round Semantic Kernel Capabilities. A process isn’t an agent, as it’s a defined workflow and there is no self-orchestration. Nevertheless, a step can comprise an agent if it has well-defined inputs and outputs. Processes can reap the benefits of widespread patterns, and there’s no purpose to have features function sequentially—they will run asynchronously in parallel, permitting you to have flows that fan out or that rely upon a number of inputs.

The 2 platforms converge of their use of Orleans, which ensures they’ve comparable approaches to working in event-driven environments. This is a vital basis, as Orleans’ transfer from being a Microsoft Analysis undertaking to being the foundational distributed computing structure for contemporary .NET has been key to wider uptake.

Utilizing AutoGen as a part of its agent tooling will assist ship higher help for multi-agent operations in Semantic Kernel. Because it’s been a analysis undertaking, there’s nonetheless some work essential to carry the 2 platforms collectively, with AutoGen supporting each .NET and Python, very similar to Semantic Kernel.

Definitely AutoGen simplifies the method of constructing brokers, with a no-code GUI and support for a variety of different LLMs equivalent to OpenAI (and Azure OpenAI). There’s additionally help for Ollama, Azure Foundry-hosted fashions, Google Gemini, and a Semantic Kernel adapter that allows you to use Semantic Kernel’s mannequin shoppers.

Getting began with AutoGen

Getting began with AutoGen requires the core AutoGen application and a model client. As soon as put in, you possibly can construct a easy agent with a handful of strains of code. Issues get attention-grabbing whenever you construct a multi-agent software or, as AutoGen calls it, a crew. Groups are introduced collectively in a gaggle chat the place customers give brokers duties. It comes with prebuilt brokers that can be utilized as constructing blocks, equivalent to a consumer proxy, an internet surfer, or an assistant.

You may shortly add your personal extensions to customise actions throughout the AutoGen layered framework. This gives particular roles for parts of an agent, beginning with the core API that gives instruments for occasion dealing with and messaging, providing you with an asynchronous hub for agent operations. Above that’s the AgentChat API. That is designed that will help you shortly construct brokers utilizing prebuilt parts and your personal code, in addition to instruments for dealing with directions and prompts. Lastly, the Extensions API is the place you possibly can add help for each new LLMs and your personal code.

A lot of the documentation focuses on Python. Though there’s a .NET implementation of AutoGen, it’s missing documentation for key features such as AgentChat. Even so, .NET is probably going one of the best device to construct brokers that run throughout distributed methods, utilizing its help for .NET Aspire and, via Aspire, frameworks like Dapr.

Constructing brokers in AutoGen Studio

AutoGen Studio is perhaps the most interesting part and would work properly as a part of the Semantic Kernel extension for Visual Studio Code. It installs as a neighborhood internet software and gives a spot to assemble groups of brokers and extensions, with the purpose of setting up a multi-agent software while not having to put in writing any extra code (although you need to use it to edit generated-configuration JSON). It builds on high of AutoGen’s AgentChat service.

Purposes are constructed by dragging parts onto the AutoGen Studio canvas and including termination circumstances. This final choice is necessary: That is how an agent “is aware of” it has accomplished a activity and must ship outcomes to both a consumer or a calling operate. Brokers may be additional configured by including fashions and extensions, for instance, utilizing an extension to ship a RAG question towards enterprise information. A number of mannequin help helps you select an appropriate AI mannequin for an agent, maybe one which’s been fine-tuned or that provides multi-model actions so you possibly can work with pictures and audio in addition to textual content prompts. Nodes in a crew may be edited so as to add parameters the place crucial.

Below the hood, AutoGen is a declarative agent improvement surroundings, with JSON description of the varied parts that go into making an agent. You may change to a JSON view to make adjustments and even convert AutoGen AgentChat Python to JSON and edit it in Studio. To simplify constructing new purposes, it provides a gallery the place brokers and different parts may be shared with different customers. When you’ve constructed an agent, you possibly can consider it inside Studio’s playground earlier than constructing it into a bigger course of.

Utilizing declarative programming techniques to construct agent groups is smart; typically the information wanted to assemble parts of a workflow or enterprise course of is embedded within the course of itself as information passes from employee to employee. If we’re to construct AI-based brokers to automate parts of these processes, who higher to design these duties than the individuals who know precisely what must be achieved?

There’s quite a bit but to return for Semantic Kernel in 2025. Now that we’re popping out of the experimental section of enterprise AI the place we used chatbots to discover ways to construct efficient prompts, it’s time to make use of these classes to construct workflow instruments extra suited to the multi-channel, multi-event processes that kind the spine of our companies. Semantic Kernel is beginning to step out into the enterprise IT world. It’ll be attention-grabbing to observe the way it and AutoGen reap the benefits of the abilities and information that exist throughout our organizations, past IT and improvement groups.