Semantic Kernel

Semantic Kernel is Microsoft's open-source SDK for building agents, and its distinguishing choice is to treat an AI model as one more service inside an ordinary application rather than as the application's centre. If you already write C#, Java or Python with dependency injection and typed interfaces, it is the framework that met you where you were.

What the pieces are
Four parts carry most of the design:

• The kernel: a container holding model connectors, services, and the registered functions the model is permitted to call.
• Plugins: groups of callable functions, where a plain annotated method becomes a tool and its schema is generated from the signature and doc comments.
• Planners and the agent framework: the loop that decides which functions to invoke, including multi-agent handoff patterns.
• Memory connectors: a uniform interface over vector stores for retrieval across steps.

Why it matters
The appeal is enterprise fit. Tool schemas derive from real typed code, so the compiler catches a signature change that would otherwise surface as a runtime tool-call failure, and the SDK slots into existing .NET and Java services where LangChain's Python-first assumptions do not reach. The trade is ecosystem size: fewer community integrations and fewer worked examples than LangChain or LangGraph, and a planner layer that has been reworked more than once as the field settled on native function calling.