Semantic Kernel (Microsoft) Platform Research

1. Introduction

1.1 Platform Positioning

Semantic Kernel (SK) is an enterprise-grade AI SDK released by Microsoft in 2023-03 under the MIT license, positioned as a “pluggable kernel that integrates LLMs with traditional code.” Its target users are .NET / Java / Python engineering teams that need to embed AI capabilities into existing enterprise applications.

By the unified framing of the project parameter card, Semantic Kernel is the enterprise-grade representative of the second, “tools & orchestration era”. Its historical value lies in being the first to bring enterprise software engineering practices such as dependency injection, type safety, the plugin model, telemetry, and the filter pipeline into LLM application development, during the window of transition from “Prompt engineering” to “Agent engineering”.

But by 2026, Semantic Kernel’s positioning has shifted to a “legacy but supported” track — it is the foundational layer for understanding Microsoft Agent Framework (MAF), but it should not be the starting point for new projects.

1.2 Basic Information

Version numbers are as of the 2026-09 retrieval date.

1.3 Clarifying Maintenance Mode Status

The exact meaning of “maintenance mode” needs to be well understood when selecting a framework:

This means: production systems already running on SK do not need a panicked migration — they can keep running reliably until at least 2027-04. But no new feature requests will be fulfilled on SK.

From late 2025 to early 2026, Microsoft launched a unification strategy: both SK and AutoGen entered maintenance mode, and their capabilities converged into Microsoft Agent Framework.

1.4 Convergence with AutoGen: Microsoft Agent Framework

Convergence timeline (see section 1.3 of 09-autogen.md):

How each predecessor contributes its legacy to MAF:

Hence a frequently quoted judgment: “Learning Semantic Kernel today is learning the foundation layer of Agent Framework.” This is accurate from an engineering standpoint — the Kernel, plugins, connectors, and filters all survive within MAF; only the Kernel has been demoted from a “god object” to optional.

Key API migration mapping:

2. Glossary

3. Feature Description

3.1 Pluggable Kernel

SK’s core organizing principle is plugins: uniformly encapsulating LLM capabilities and traditional code as “named function containers,” then dispatched centrally by the Kernel. Plugins can be created in four ways:

This design was quite ahead of its time in 2023 — it turned “tools” from a special mechanism into first-class citizens isomorphic with business code.

3.2 AI Service Connectors

Covers Azure OpenAI / OpenAI / Amazon Bedrock, with a capability surface including chat, text generation, embeddings, images, and audio.

Important trait: SK does not require Azure — it can run on any cloud or on-premises. This is often misunderstood during enterprise selection: SK is a Microsoft product, but not an Azure-only product.

3.3 Planner-Driven Orchestration

SK’s orchestration approach is the Planner: the LLM dynamically composes sequences of function calls. This is in sharp contrast to LangGraph’s explicit state graph:

This contrast is a classic embodiment of the “flexibility ↔ predictability” tension, and in 2023 SK stood on the flexibility side. The 2026 industry consensus is that, in production scenarios requiring reliability guarantees, explicit graphs outperform dynamic Planners — which is also the practical reason MAF ultimately adopted graph workflows.

3.4 Enterprise-Grade Capabilities

3.5 Multi-Language SDK and Feature Parity

SK provides official SDKs in three languages: C# (primary), Python, Java. This is a unique advantage over most Python-first frameworks, especially significant for .NET and Java enterprise stacks.

It should be noted that no detailed comparison of feature parity across the three languages has been found: the specific gaps among language versions in plugin ecosystems, the Memory abstraction, and Planner implementations lack an official comparison table. This is an explicit information gap.

4. Platform Architecture

4.1 Layered Architecture

图 4-1｜Semantic Kernel 五层架构：以 Kernel 为中心的插件化内核

数据来源：基于本文分析绘制的示意图。

Client / Application（Web / Console / Teams Bot / 后台服务）
   ↑
Agent 层（ChatCompletionAgent / AzureAIAgent / OpenAIAssistantAgent）
   ↑
Runtime 层（Session 管理 / Middleware 管线 / Memory & Context / 执行循环）
   ↑
Tool / Integration 层（Function Tools / MCP 客户端 / 外部 API / 数据库）
   ↑
Model Provider 层（Azure OpenAI / OpenAI / Anthropic / Gemini / Ollama）

Note: these five layers are the form after MAF inherited and evolved them. SK’s original form centered on the Kernel, with the Agent layer added later and always in an experimental state.

4.2 Kernel-Centric Model

The core trait of SK’s original form is: everything must connect through a built Kernel. The Kernel bears multiple responsibilities — DI container, configuration surface, plugin registry, and service selector — which is why it is called a “god object”.

The cost of this design is a lot of boilerplate — even to run the simplest agent, you must first build a Kernel, register services, and add plugins. This is exactly the friction MAF set out to eliminate.

4.3 MAF Inheritance and Evolution

5. Harness Design

5.1 Six-Layer Capability Overview

5.2 L1 Context Engineering Layer

Rating: Medium.

Mechanisms provided: Prompt template language (Handlebars, etc.) for structured prompt assembly; text-search plugins for RAG.

Missing mechanisms: no native automatic context compression / compaction. This is understandable against the 2023 technical backdrop (the context window was the main constraint and agent task chains were short), but in 2026 long-horizon task scenarios this is a hard gap.

Engineering implication: long conversations on SK must implement summarization or window truncation themselves, and this cannot coordinate with the Planner’s intermediate steps — because the Planner’s path is inherently dynamic, you don’t know where to truncate.

5.3 L2 Tools & Execution Layer

Rating: Medium-strong.

• Native plugins: annotated with [KernelFunction] / @kernel_function.
• OpenAPI spec import: automatically generates plugins from API definitions.
• Vector-store connectors: Azure AI Search, Elasticsearch, etc., exposable as plugins for retrieval and semantic search.

Key gap: MCP support on the SK side is incomplete. MCP has become the de facto standard for tool interoperability in 2026, and since SK is in maintenance mode it will not fill this capability in — teams that need MCP must move to MAF. This is the platform’s most decisive technological obsolescence factor.

5.4 L3 Orchestration & Control Layer

Rating: Medium.

• Planner dynamic composition: Stepwise / Handlebars / custom.
• Process Framework: structured business workflows.
• Agent orchestration has always remained experimental — this is SK’s most fatal positioning problem. It provides agent types (ChatCompletionAgent, etc.), but has never organized these agent types into a production-grade multi-agent orchestration system. Only MAF added production-grade orchestration and persistent checkpoints.

Tension 1 (flexibility ↔ predictability): SK’s Planner is a purely flexibility-oriented design — the LLM dynamically decides which functions to call and in what order. This was innovative in 2023, but by 2026 it is regarded as a reliability anti-pattern. The industry has already voted with its feet: MAF replaced the Planner with the “five modes + explicit graphs.”

5.5 L4 Memory & State Layer

Rating: Medium.

• Thread: conversation thread abstraction (but the corresponding subclass must be instantiated manually).
• Memory abstraction: three layers — semantic memory / long-term memory / entity memory.
• Vector-store connectors: Azure Cognitive Search, Chroma, Qdrant, pgvector.

No persistent checkpoint: long-running flows cannot be resumed across processes. This is a common shortcoming of the second-generation frameworks, and it is where LangGraph builds its advantage through the Checkpointer.

5.6 L5 Evaluation & Observability Layer

Rating: Medium.

• OpenTelemetry telemetry and diagnostics: standardized observability output; this is one of SK’s most solid enterprise-side capabilities.
• Filters: provide interception points usable for instrumentation and quality detection.

No built-in evaluation suite. SK provides the infrastructure for “what you can observe,” but not the methodology or tooling for “how you should judge quality.”

5.7 L6 Governance & Security Layer

Rating: Medium-strong.

This is the most forward-looking part of SK relative to its era, and the reason MAF fully inherited it:

The Filters mechanism deserves special emphasis: back in 2023 it introduced the pattern of “inserting enterprise policy checks before and after every LLM/function call.” This pattern later became industry standard — CrewAI’s runtime hooks, MAF’s Middleware, and Alibaba Cloud Bailian’s Middleware are all different implementations of the same idea.

On cost guardrails: no task-level budget-cap mechanism was found for SK. Filters could theoretically be used for counting and rate limiting (would need to be implemented yourself), but there is no out-of-the-box capability.

5.8 Projection of the Three Inherent Tensions onto Semantic Kernel

A historical observation worth recording: SK was ahead on “governance” but behind on “openness.” It used Filters to push governance down to the call level (finer-grained than most 2026 platforms), yet because it lacked MCP it was marginalized in tool interoperability. This shows that Harness’s competitiveness is not single-point depth, but the synchronized evolution of all six layers — if any layer falls behind, it drags down the whole.

6. Real-World Cases

No enterprise deployment case with quantified impact data was found. The retrieved materials are all framework comparisons and migration guides.

Qualitative points that are available:

• Since its 2023-03 release and until entering maintenance mode, Semantic Kernel accumulated about three years of enterprise field experience.
• SK and AutoGen together accumulated more than 75,000 GitHub stars, and this engineering asset has been inherited by MAF.
• SK’s enterprise-grade capabilities (type safety, filter pipeline, telemetry, multi-language) have a real deployment base in .NET and Java enterprise stacks, but no publicly citable customer names or impact metrics were found.

Honest note: all of the above are qualitative descriptions and do not constitute evidence of deployment impact.

7. Summary

7.1 Strengths

• Stable, production-ready: long-term support until at least 2027-04; existing workloads can run with confidence.
• Enterprise-grade type safety and telemetry: OpenTelemetry native; Filters provide call-level interception.
• Not tightly bound to Azure: can run on any cloud or on-premises.
• Multi-language SDK (C# / Python / Java): highly significant for .NET and Java enterprise stacks, and rare in the Python-dominated ecosystem.
• Learning it equals learning MAF’s foundation layer: Kernel, plugins, connectors, and filters all survive within MAF, so the learning investment is not fully sunk.
• MIT license, with no commercial restrictions.

7.2 Weaknesses

• Maintenance mode, no new features: no new requirements will be fulfilled on SK.
• Agent orchestration remains experimental: never formed a production-grade multi-agent orchestration system.
• No native MCP / A2A support: marginalized under the 2026 tool-interoperability standard.
• No persistent checkpoint: long-running flows cannot be resumed across processes.
• Must build through the Kernel god object: a lot of boilerplate.
• No native context compression: long-horizon tasks must manage context budgets themselves.
• Reliability issues with Planner dynamic orchestration: paths are not enumerable and are hard to debug.
• No public quantified deployment data was found.

7.3 Applicability Boundaries

New projects should not choose Semantic Kernel. They should use Microsoft Agent Framework directly.

Scenarios where it can still be retained:

• Existing SK production workloads with no new-feature requirements — they can run reliably for more than a year until 2027-04.
• Simple integrations needing only “LLM + plugin + telemetry,” not involving multi-agent orchestration, where the team is already familiar with SK.
• As a path for learning MAF’s foundation layer (but this does not constitute a production selection reason).

7.4 Selection & Migration Recommendations

Recommended strategy: “Use the new framework for new agents, and migrate existing agents opportunistically when touched.”

Technical-feasibility assessment of migration: since MAF retains the Kernel, plugin model, connectors, and filters, and has released an official migration guide (from the 2026-02 RC period), the SK → MAF migration is of the type “concepts map directly, APIs need rewriting” rather than a from-scratch rebuild. The main change points:

1. Multiple agent types → unified ChatClientAgent.
2. Manual Thread management → agent.CreateSessionAsync().
3. Kernel mandatory construction → optional.
4. Filters → Middleware (same concept, different API).
5. Planner → graph workflows / the five orchestration modes (this is the only part that needs redesigning).

One-sentence conclusion: Semantic Kernel is an important foundational pioneer of enterprise-grade LLM SDKs, and its Filters and plugin model remain industry-benchmark designs to this day. But by 2026 it is already on a “legacy but supported” track — it is the key to understanding MAF, not the road to MAF.

Information Gap Statement

1. Enterprise deployment cases and quantified impact data: after three rounds of retrieval, no enterprise deployment case with quantified impact data was found. Everything retrieved is framework comparison and migration guides.
2. Feature parity across languages: for the specific gaps among the C# / Python / Java SDKs in plugin ecosystems, the Memory abstraction, and Planner implementations, no official detailed comparison table was found.
3. Exact end date of maintenance mode: only “supported for at least until 2027-04” (i.e., at least one year after MAF GA) is known; the exact end-of-support date is not public.
4. Exact boundary of MCP support: the report records that “MCP support on the SK side is incomplete,” but no official statement on “to what extent it is supported” was found, marked [To be verified].
5. Task-level cost guardrails: no budget-cap mechanism was found for SK, marked [To be verified].
6. Version number is time-sensitive: Microsoft.SemanticKernel 1.77.0 (2026-05-28) is the version at the 2026-09 retrieval point.
7. Transparency / AI-generated content labeling: no public materials were found from Semantic Kernel’s official channels on transparency statements or AI-generated content labeling policy.
8. Complete SK-to-MAF API migration comparison table: the retrieved migration guides cover the main change points (agent types, Thread, Kernel, Filters), but lack a complete per-API comparison, marked [To be verified].

8. References

1. Semantic Kernel: Enterprise Agent Framework — Agent DevPro. https://agentdevpro.com/frameworks/semantic-kernel
2. Microsoft's Quiet Shift: Why Semantic Kernel Is Moving to Maintenance Mode — FrontierNews. https://www.frontiernews.ai/news/article/microsofts-quiet-shift-why-semantic-kernel-is-movi-23683b05
3. Migrate a Semantic Kernel App to Microsoft Agent Framework 1.0 — Start Debugging. https://startdebugging.net/2026/07/migrate-a-semantic-kernel-app-to-microsoft-agent-framework-1-0
4. Microsoft Agent Framework vs Semantic Kernel for a Greenfield .NET Agent — Start Debugging. https://startdebugging.net/2026/06/microsoft-agent-framework-vs-semantic-kernel-for-a-greenfield-net-agent
5. Microsoft Agent Framework 1.0 Ends the Semantic Kernel Era — SourceFeed. https://sourcefeed.dev/a/microsoft-agent-framework-10-ends-the-semantic-kernel-era
6. Microsoft Agent Framework — Microsoft Learn official documentation. https://learn.microsoft.com/en-us/agent-framework/overview/
7. Microsoft Agent Framework 1.0: .NET and Python 2026 — DigitalApplied. https://www.digitalapplied.com/blog/microsoft-agent-framework-1-0-dotnet-python-guide
8. Microsoft Agent Framework Developer Guide 2026 — Developer's Digest. https://www.developersdigest.tech/blog/microsoft-agent-framework-developer-guide-2026
9. Microsoft Agent Framework at BUILD 2026 — Microsoft official. https://aka.ms/Build2026MicrosoftAgentFramework