← Back to concept explainers

Concept Fable

Semantic Kernel

Conclusion first:Semantic Kernel is Microsoft's open-source AI orchestration SDK, positioned to integrate large language models, plugins, planners, memory and native application code into enterprise applications. Its core abstractions include: Kernel (the orchestration core), Plugins (capability encapsulation, which can contain Semantic Functions and Native Functions), Planner (decomposing user intent into multi-step execution plans), and Memory (semantic memory and retrieval via vector stores). Semantic Kernel's biggest feature is enterprise-grade embedding—it does not run as a standalone service, but is embedded into existing .NET / Python / Java applications, reusing the enterprise's existing authentication, logging, configuration and deployment systems. In RAG and Agent scenarios, Semantic Kernel's value is injecting AI capabilities into existing business systems as an SDK, rather than standing up a separate standalone service.

Semantic Kernel · Enterprise AI orchestration SDK

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The central dispatch desk does not need to learn every workshop's craft

In the royal city there was a crafts dispatch hall, where there were both scribes who could write letters and accountants who checked ledgers, draftsmen who drew plans, and couriers who ran errands. In the past, when a client wanted to handle a complex matter, they had to relay messages between rooms themselves, wearing out their legs and often still missing materials.

The new steward registered each craft as a callable work badge: this one can audit accounts, and needs which materials; that one can draw diagrams, and delivers what format; the courier can deliver letters, but needs an address and a seal. If a scribe wants to complete a business letter, he can call on people by badge to supplement evidence, draw sketches and send back receipts.

The first version of the dispatch hall was chaotic. The scribe called whoever came to mind, the accountant's requested materials were incomplete, and the courier got half an address. Customers only felt that "the hall is very smart," but when it came to actually getting things done, it got stuck on details.

The manager then wrote the process: first understand the request, then choose the craft, then pass in the materials, then connect the result to the next step. Each work badge must also state how to report failure, who needs to approve, and who receives the result. Registering crafts is not enough; crafts must also collaborate in order.

Later, common tasks were compiled into small kits. Writing quotes, checking inventory, drafting replies, and scheduling deliveries could each be combined and swapped. What the customer saw was one smooth service, but behind it were many crafts connected by agreement.

Once, when the accounting office was temporarily closed, the dispatch hall did not pretend to finish; instead, it paused the task at the audit step and notified the scribe to first draft the part without amounts. When the accounting office reopened, the process continued from there. Only then did the manager feel reassured: dispatch is not about hiding risk, but about making every step visible.

The value of the dispatch hall lies in organizing eloquent scribes and executable work badges. A single smart person is not enough; only by stringing intent, skills, memory, and steps into a reliable process can complex work run.

Later, the dispatch hall was expanded, adding accounting, goods inspection, and seal-carving crafts. The manager did not let the scribes call on them at will, but first added work badges and processes. The larger the hall, the more unified conventions are needed; otherwise, the more crafts there are, the more likely a customer's errand is to break down halfway. Later, customers could not see the badges behind the scenes, but they could feel that errands had fewer breakpoints and less repeated explanation.

Reveal

This story is about: Semantic Kernel

Semantic Kernel is Microsoft's open-source AI orchestration SDK, positioned to integrate large language models, plugins, planners, memory and native application code into enterprise applications. Its core abstractions include: Kernel (the orchestration core), Plugins (capability encapsulation, which can contain Semantic Functions and Native Functions), Planner (decomposing user intent into multi-step execution plans), and Memory (semantic memory and retrieval via vector stores). Semantic Kernel's biggest feature is enterprise-grade embedding—it does not run as a standalone service, but is embedded into existing .NET / Python / Java applications, reusing the enterprise's existing authentication, logging, configuration and deployment systems. In RAG and Agent scenarios, Semantic Kernel's value is injecting AI capabilities into existing business systems as an SDK, rather than standing up a separate standalone service.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Each branch workshop understands the capabilities of all branch workshops: application code hardcodes all AI call paths, resulting in poor scalability
  • The owner translating and dispatching alone: using a single LLM prompt to handle all task routing, with no structured orchestration
  • The central control console doesn't care about process details: Semantic Kernel's Kernel abstraction only handles task orchestration
  • Interface specification declaration: the function signatures and descriptions of Plugins, so the Kernel knows what each capability can do
  • Cross-task memory book: Semantic Kernel Memory, using vector storage to carry conversation history, user preferences, and domain knowledge
  • Split into steps for dispatch: the Planner decomposes complex intent into a multi-step execution chain
  • Standard slot registration: plugin architecture, new capabilities connect as Plugins without changing the Kernel core
  • Each workshop does only what it does best: separation of concerns—LLM handles reasoning, Plugin handles execution, Kernel handles orchestration

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

pluginplannerkernelfunctionmemoryMicrosoft Copilot stack
京ICP备2026014272号