Semantic Kernel

Semantic Kernel is an open-source SDK from Microsoft for building AI agents and integrating LLMs into enterprise applications in C#, Python, and Java. It acts as middleware between your existing code and AI models, letting you define functions and plugins that an LLM can call automatically to complete tasks. Microsoft uses Semantic Kernel internally across its own enterprise products, and it has version 1.0 stability commitments across all three supported languages, with built-in telemetry, filters, and security hooks designed for production deployment. It is now the foundation of Microsoft Agent Framework, Microsoft’s enterprise-grade successor for multi-agent orchestration.

How Semantic Kernel works

The key idea is the Kernel: a dependency injection container that holds all your AI services, plugins, and memory. When a user sends a request, the Kernel selects the right service, assembles the prompt, sends it to the LLM, and routes any function calls back to your code. Here is how the components fit together:

• Kernel: The central engine. It manages all registered AI services and plugins, monitors each step through event hooks, and is the single place where you configure and observe your agent.
• Plugins: Collections of functions you expose to the LLM. A plugin can wrap native code (a Python function, a C# method), a semantic prompt template, an OpenAPI specification, or an MCP server. The LLM reads the plugin descriptions and decides when to call them.
• Function calling: The mechanism by which the LLM requests a specific plugin function during a conversation. Semantic Kernel intercepts the request, executes the matching function, and returns the result so the LLM can continue reasoning.
• Memory: A vector-based memory layer that lets agents store and retrieve information across conversations. It integrates with Azure AI Search, Pinecone, Chroma, Qdrant, and other vector stores to support RAG.
• Agents: Higher-level components built on the Kernel. A ChatCompletionAgent wraps a kernel, a set of plugins, and an instruction prompt into a single callable agent. Multiple agents can be orchestrated together in collaborative workflows.
• Process Framework: A workflow layer for modelling long-running, stateful business processes with explicit steps, branching, and human-in-the-loop checkpoints.
• Filters: Middleware hooks that run before and after function calls, enabling logging, content moderation, responsible AI checks, and observability at every step.

What you can build with Semantic Kernel

• Enterprise CRM assistant: Wrap your CRM API as a plugin so an agent can look up account details, update records, and summarise deals directly from a chat interface, with your existing auth and permission layer intact.
• Automated report writer: Build a multi-agent system where one agent gathers data from internal APIs through plugins, a second synthesises findings, and a third formats and writes the final report.
• Smart home control agent: Register device control functions as plugins so the agent can interpret natural language commands like “dim the lights in the office” and execute the correct API call without custom routing logic.
• Supply chain optimisation assistant: Connect database query functions and calculation logic as plugins, then let the agent break a complex business question into sub-steps, call each plugin in the right order, and return a reasoned answer.
• Long-running business process: Use the Process Framework to model a multi-step approval workflow with explicit state management, where agents handle each step and humans review and approve at defined checkpoints.
• Multimodal customer support bot: Use Semantic Kernel’s multimodal support to build a support agent that processes both text and images from customer submissions, routing each to the right plugin for analysis and response.