Semantic Kernel

Semantic Kernel is Microsoft's open-source orchestration SDK for building AI agents and LLM-powered features inside production applications. It gives you a Kernel object that binds prompts, plugins (native code or OpenAPI tools), memory/vector stores, and model connectors together, plus higher-level Agent and Process frameworks for multi-step, multi-agent workflows. SDKs ship for C#, Python, and Java, which makes it one of the few serious agent frameworks that treats the .NET/JVM enterprise stack as a first-class citizen rather than a Python afterthought.

It's aimed at engineers who need to embed LLMs into existing enterprise systems and want Microsoft-grade observability, filters, and security hooks rather than a bare LangChain-style toolkit. Semantic Kernel itself is free and MIT-licensed; the cost is whatever model API you point it at (Azure OpenAI, OpenAI, Hugging Face, local models, etc.). Compared to LangChain it's leaner and more opinionated about dependency injection and typed contracts, and compared to AutoGen it's more focused on shipping into real apps than on research-y agent choreography.

Strong Azure integration is both the selling point and the trap: connectors for Azure AI Search, Azure OpenAI, and Cosmos DB are excellent, but non-Microsoft ecosystems can feel like second-class citizens. The Agent Framework and Process Framework are still evolving, and API churn between versions has bitten early adopters. For teams already committed to .NET or Azure it's close to a default choice.