// the find
microsoft/semantic-kernel
Integrate cutting-edge LLM technology quickly and easily into your apps
Semantic Kernel is Microsoft's SDK for building LLM-powered agents and multi-agent systems in C#, Python, and Java. It handles the plumbing: connecting to model providers, wiring up tool/function calling, managing conversation history, and orchestrating multi-agent workflows. The README's opening notice is the real story — Microsoft has declared this project superseded by their new Agent Framework, and is actively pushing users to migrate.
The plugin system is genuinely well-designed: decorating a C# method with [KernelFunction] and a description is all you need to expose it as an LLM-callable tool, and it just works. Multi-provider support is broad and not superficial — OpenAI, Azure OpenAI, Hugging Face, Ollama, ONNX, and others have real connector implementations, not just API key passthrough. The vector store abstraction covers a wide range of backends (Azure AI Search, Elasticsearch, Chroma, and more) with a consistent interface, which matters when you want to swap providers later. The architecture decision records in docs/decisions/ are unusually thorough for a Microsoft OSS project — 73 ADRs documenting why things are the way they are.
The project is effectively deprecated: the README leads with a migration notice to Microsoft Agent Framework 1.0, so any investment here is investment in something Microsoft has already moved on from. The abstraction layers add real cognitive overhead — KernelArguments, PromptExecutionSettings, FunctionChoiceBehavior — you spend time understanding the framework's model of the world before you can do anything useful. Error handling at the agent orchestration level is thin; if a tool call fails or a model returns something unexpected mid-workflow, the failure modes are hard to reason about. The Java implementation lags behind C# and Python significantly, so cross-language parity claims should be treated skeptically.