Semantic Kernel let a model choose which function to call

The conversation

In late February 2023, Microsoft opened the Semantic Kernel project on GitHub, and on 17 March 2023 published a companion devblogs post describing it as "a lightweight SDK that lets you mix conventional programming languages, like C# and Python, with the latest in Large Language Model (LLM) AI 'prompts' with prompt templating, chaining, and planning capabilities." The pitch was integration: rather than choosing between a hand-written function and a model-generated one, a developer could register both as skills the kernel could call. The post describes a Planner component that takes "a user's 'ask'" and works out "the Skills, Memories, and Connectors needed to achieve their goal," which Microsoft calls "a goal-oriented approach to programming."

What the documents show

Retrieved 16 September 2026, Microsoft's current Semantic Kernel overview describes the same underlying mechanism in production terms: "Semantic Kernel combines prompts with existing APIs to perform actions. By describing your existing code to AI models, they'll be called to address requests," and "Semantic Kernel is the middleware translating the model's request to a function call and passes the results back to the model," a model-driven selection process, not a fixed, hand-coded call path. The project's own GitHub repository, also retrieved 16 September 2026, carries a notice absent in 2023: "Semantic Kernel is now Microsoft Agent Framework," describing that successor as the "enterprise-ready successor to Semantic Kernel," now "available at version 1.0."

The system boundary

Across both the 2023 and 2026 documents, Semantic Kernel's plugins are functions a developer writes and registers, with the model choosing among them by matching a request to a function's stated purpose rather than by direct access to a developer's code. The documentation names this middleware role explicitly: it is the layer that turns a model's request into an actual function call, standing between what a model decides to request and any code that executes. Which functions exist to be chosen, and what those functions are permitted to do, remains a decision the developer makes when registering a skill, not one the planner can expand on its own.

Where it fails

Because planner-driven selection is model-based rather than deterministic, the same registered plugin set can be assembled into different call sequences for similarly worded requests, and neither the 2023 post nor the current overview states that a planner's chosen sequence is checked against a specification before it runs. The GitHub repository's own rename notice is itself evidence to watch for: a team building against tutorials written for Semantic Kernel by name should confirm whether that code path still receives updates under that name or has moved to Microsoft Agent Framework.

• Is the planner's chosen sequence of function calls reviewed before it executes, or only after?
• What can each registered plugin function actually do, and is that scope documented anywhere a reviewer can check it?
• Is the project version being used still Semantic Kernel, or should it be migrated to its stated successor?

From a February 2023 GitHub repository to a 2026 rename notice, Microsoft's own documents describe a middleware layer for connecting prompts to code, now explicitly in transition to a differently named successor project.