Exploring AutoGen, Semantic Kernel and the New Microsoft Agent Framework
Notes by SDG ·
The rapid rise of large‑language models (LLMs) has given developers new ways to build intelligent applications. Simple chatbots have evolved into autonomous AI agents—components that can reason about a goal, call tools or APIs, maintain context and even collaborate with other agentsdevblogs.microsoft.com
. Microsoft’s open‑source projects AutoGen and Semantic Kernel were created to explore two sides of this spectrum: experimental multi‑agent orchestration on one side and enterprise‑ready LLM integration on the otherdevblogs.microsoft.com
. In October 2025 the two projects were unified into the Microsoft Agent Framework, an open‑source SDK and runtime that brings research‑grade orchestration together with production‑ready stabilitydevblogs.microsoft.com
. This write‑up starts from first principles—explaining what AutoGen and Semantic Kernel are, illustrating how they combine in the new framework, outlining practical use‑cases, comparing them with other agentic frameworks and explaining why developers might choose Microsoft’s solution.
What is AutoGen?
AutoGen is an open‑source programming framework for building AI agents that can converse with each other and with humansmicrosoft.github.io
. Developed by Microsoft Research, it treats an agent system as an asynchronous conversation among specialized agents; each agent can be a ChatGPT‑style assistant or a tool executor, and conversations evolve via messageslangfuse.com
. Key features include:
- Multi‑agent conversation framework: AutoGen simplifies the orchestration and automation of complex LLM workflows by enabling multiple agents to chat and collaboratemicrosoft.github.io
- . These conversations can be dynamic, concurrent and event‑drivenlangfuse.com
- .
- Diverse conversation patterns: The framework supports a range of conversation patterns, from single‑agent interactions to group chats and tool‑calling patternsmicrosoft.github.io
- .
- Tool and code execution: Agents can execute code locally or inside Docker containers, call external tools or open‑source functions and incorporate human oversight where neededmicrosoft.github.io
- .
- Research‑driven: The project grew out of collaborative research from Microsoft, Penn State University and the University of Washingtonmicrosoft.github.io
- and continues to incubate cutting‑edge multi‑agent patterns.
Where AutoGen Shines
AutoGen is ideal when your application revolves around multi‑turn conversations and real‑time interactions. Examples include:
- Collaborative research: A planning agent delegates tasks to researcher and summarizer agents, each using external tools or web searches. The asynchronous conversation continues until the planner decides the task is complete.
- Coding assistants with feedback: A developer prompts a code‑assistant agent. The assistant writes code, sends it to a user‑proxy agent that executes the code, and receives the output (or errors) to iterate on the solution. The quickstart example in AutoGen’s documentation shows how a user proxy can ask an assistant to plot stock prices; the assistant writes Python code and the user proxy executes itmicrosoft.github.io
- .
- Customer‑service teams: Several specialized agents (e.g., a sentiment‑analysis agent, a policy‑compliance agent and a response‑generation agent) converse to craft responses that comply with regulations while addressing customer concerns.
AutoGen’s asynchronous chat paradigm reduces blocking, making it well suited for workflows that require waiting on external APIs or human inputlangfuse.com
. However, because it emphasizes dynamic conversation and experiments, AutoGen on its own lacks enterprise‑grade observability, long‑running durability and compliance hooksdevblogs.microsoft.com
.
What is Semantic Kernel?
Semantic Kernel (SK) is a lightweight, open‑source SDK that integrates LLMs into existing codebases for C#, Python and Javalearn.microsoft.com
. It acts as a middleware layer between AI models and application code; developers can describe their functions via prompts or OpenAPI specs, and the kernel orchestrates calls to these functionslearn.microsoft.com
. The project emphasizes enterprise readiness and extensibility. Highlights include:
- Model‑agnostic connectors: SK provides connectors for OpenAI, Azure OpenAI, Hugging Face and Nvidia modelsraw.githubusercontent.com
- . New models can be swapped in without rewriting application codelearn.microsoft.com
- .
- Plugins and skills: Developers can register existing code as plugins (functions) or prompt templates. These plugins, plus built‑in connectors for Microsoft Graph, SharePoint, Elastic, Redis and other systems, allow SK to call enterprise APIs directlyraw.githubusercontent.com
- .
- Multi‑agent systems: Beyond simple chat agents, SK supports orchestrating multiple specialist agents and workflows. The README notes that SK includes an agent framework, plugin ecosystem, vector database integration and a process framework for structured workflowsraw.githubusercontent.com
- .
- Enterprise readiness: Telemetry, security hooks and stable APIs make SK suitable for production uselearn.microsoft.com
- . It supports vector memory across databases (e.g., Azure AI Search, Elasticsearch, Pinecone) and can run on local environments such as Ollama or ONNXraw.githubusercontent.com
- .
Where Semantic Kernel Excels
Semantic Kernel shines when developers need to augment existing business applications with AI. For example:
- Automating business processes: SK pairs LLM prompts with existing APIs so that, when a user request arrives, the model calls the appropriate function and returns resultslearn.microsoft.com
- . This pattern is useful for automating workflows such as processing insurance claims, approving travel expenses or managing HR tickets.
- Retrieval‑augmented generation: SK integrates with various vector stores and search services to provide long‑term memory; this is essential for answering questions over enterprise knowledge bases or summarizing large document setsraw.githubusercontent.com
- .
- Complex planning and process orchestration: Through its process framework, SK can model deterministic, long‑running business processes with checkpoints and human‑in‑the‑loop interactions. Such structured workflows are ideal for regulated industries where reliability and compliance matter.
Because of its emphasis on enterprise connectors, state management and type safety, Semantic Kernel has been widely adopted by Fortune 500 companies and is considered a strong choice for mission‑critical AI integrationslearn.microsoft.com
.
Microsoft Agent Framework: Where AutoGen and Semantic Kernel Meet
Motivation
Despite the excitement around agentic AI, developers found that existing frameworks were fragmented—AutoGen offered innovative multi‑agent orchestration but lacked enterprise maturity, while Semantic Kernel provided enterprise stability but not the latest research patternsdevblogs.microsoft.com
. Developers asked Microsoft to unify these strengthsdevblogs.microsoft.com
. In response, on 1 October 2025 Microsoft announced the Microsoft Agent Framework (MAF), an open‑source SDK and runtime that combines AutoGen’s orchestration patterns with Semantic Kernel’s enterprise‑ready foundationsdevblogs.microsoft.com
.
Core Pillars
The Microsoft Agent Framework introduces four pillars to bridge research and production:
- Open Standards & Interoperability: The framework supports the Model Context Protocol (MCP), enabling agents to discover and invoke external tools over a standard protocol; Agent‑to‑Agent (A2A) messaging for collaborating across runtimes; and an OpenAPI‑first design, allowing any REST API with an OpenAPI specification to be imported as a tool without custom wrappersdevblogs.microsoft.com
- .
- Pipeline for Research: Experimental orchestration patterns from AutoGen—such as sequential, concurrent, group chat, handoff and Magentic (dynamic task ledger) workflows—are available with durability and governancedevblogs.microsoft.com
- . An extension package lets innovators try out cutting‑edge patterns while keeping stable features separatedevblogs.microsoft.com
- .
- Extensible by Design: MAF is 100 % open source with a modular architecture. It inherits connectors to enterprise systems (Azure AI Foundry, Microsoft Graph, SharePoint, Oracle, Amazon Bedrock, MongoDB and many SaaS services through Azure Logic Apps) and offers pluggable memory modules using Redis, Pinecone, Qdrant, Weaviate, Elasticsearch or Postgresdevblogs.microsoft.com
- .
- Ready for Production: Built‑in observability (OpenTelemetry), security (Entra ID), approval workflows, CI/CD integration with GitHub Actions/Azure DevOps and long‑running durability make MAF suitable for enterprise deploymentdevblogs.microsoft.com
- .
How MAF Unifies AutoGen and Semantic Kernel
Table 1 in the Azure AI Foundry announcement contrasts the older projects and the new frameworkdevblogs.microsoft.com
. Semantic Kernel focused on a stable SDK with connectors, workflows and observability; AutoGen emphasized experimental multi‑agent orchestration from research; the Microsoft Agent Framework combines these strengths into a unified SDK offering built‑in connectors and memory, standardized protocols (MCP, A2A, OpenAPI), deterministic and dynamic orchestration, and enterprise‑grade observabilitydevblogs.microsoft.com
The new framework offers a pipeline from research to production—developers can prototype local multi‑agent workflows using the same abstractions that later run in Azure AI Foundry or the Microsoft 365 Agent SDKdevblogs.microsoft.com
. It supports both Agent Orchestration (creative reasoning and decision‑making using LLMs) and Workflow Orchestration (business‑logic driven, deterministic processes)devblogs.microsoft.com
, letting teams choose between flexible collaboration and structured processes.
Practical Use‑Cases and Business Applications
MAF’s design has already attracted enterprise adopters. Visual Studio Magazine reports that early users include KPMG (automating audit testing and documentation), BMW (analyzing terabytes of vehicle telemetry), Commerzbank (piloting avatar‑driven customer support), Fujitsu (balancing human and AI collaboration in integration services) and more. These examples illustrate how MAF helps automate regulated workflows and high‑volume data analysis while maintaining observability.
Developers can get started quickly: a basic agent in Python requires a few lines of code (e.g., installing via pip install agent-framework and instantiating a ChatAgent). Sample projects include a HaikuBot that generates poems and a Hello World multi‑agent workflow connecting a writer and editor agent. The framework supports various workflow patterns—sequential, concurrent, handoff and group chat—and allows agents to call external tools via MCP servers, hosted interpreters or APIs.
MAF’s ability to integrate with existing enterprise systems means it can solve a range of business problems:
- Compliance and auditing: Agents can automatically gather evidence, perform checks against regulatory rules and produce audit logs. KPMG notes that the observability and governance features provide the controls required in regulated industries.
- Industrial analytics: BMW uses multi‑agent systems to analyze vehicle telemetry; the durability and observability of MAF allow engineers to act on insights immediately.
- Customer support and marketing: Commerzbank’s avatar‑driven support agents use MAF’s connectors and memory to provide personalized responses while ensuring compliance with bank policies.
- Knowledge management: A multi‑agent workflow could ingest unstructured documents into vector stores, classify them and answer employee questions. MAF’s retrieval support, memory abstraction and ability to combine AI and non‑AI functions make this straightforward.
The Microsoft Agent Framework stands out for several reasons:
- Unified Research and Production: It combines AutoGen’s cutting‑edge orchestration (group chat, debate, reflection, handoff patterns) with Semantic Kernel’s enterprise‑grade stability and connectorsdevblogs.microsoft.com
- devblogs.microsoft.com
- .
- Open Standards and Interoperability: MCP, A2A and OpenAPI support mean agents can call any REST API, discover tools dynamically and collaborate across runtimesdevblogs.microsoft.com
- . This cross‑framework compatibility reduces vendor lock‑in and fosters a growing ecosystem of tools.
- Enterprise‑Ready from Day One: Built‑in telemetry, observability, security and CI/CD integration ensure that prototypes can move into production without rewriting codedevblogs.microsoft.com
- . This is critical for regulated industries where compliance and audit trails matter.
- Modular and Extensible: Developers can choose their memory backend, integrate with multiple AI services and extend capabilities via connectors and pluginsdevblogs.microsoft.com
- . YAML/JSON declarative definitions allow workflows to be version‑controlled and shared across teams.
- Ecosystem Alignment: The framework aligns with the Microsoft 365 Agents SDK and Azure AI Foundry, offering a consistent runtime whether agents run locally, in Azure or within Microsoft 365 Copilotdevblogs.microsoft.com
- . Early enterprise adoption by major organizations demonstrates its viability.
Conclusion
AI agents are becoming a new layer of application logic. Microsoft’s early investments produced two complementary projects—AutoGen for exploring multi‑agent conversations and Semantic Kernel for integrating LLMs into enterprise workflows. By unifying these into the Microsoft Agent Framework, the company offers developers a single pro‑code foundation that spans research‑driven orchestration, open standards, rich connectors and enterprise‑ready operationsdevblogs.microsoft.com
. Whether you’re building a creative brainstorming assistant, automating audit workflows or orchestrating complex data‑analysis pipelines, MAF provides the building blocks to move from prototype to production without sacrificing innovation or compliance.