DEVELOPMENT TECHNOLOGIES

Microsoft Agent Framework and Semantic Kernel: Architecture and Semantic Kernel vs LangChain

How Semantic Kernel and its successor, Microsoft Agent Framework, bring LLM agents and workflows into .NET, Python, and Java backends, and how they compare with LangChain and LangGraph.
Microsoft logo

What Is Microsoft Agent Framework and Semantic Kernel?

Semantic Kernel is an open-source, model-agnostic SDK that brings large language models into C#, Python, and Java code through a central dependency injection container, plugins, and vector memory. Microsoft Agent Framework is its unified 1.0 successor, combining Semantic Kernel’s enterprise foundations with AutoGen’s multi-agent orchestration.

When architects compare Semantic Kernel vs LangChain, they are looking at two generations of Microsoft’s open-source AI orchestration stack. Both are MIT-licensed and free; you pay only for infrastructure and model usage. The abstractions are documented in the Semantic Kernel documentation and the Microsoft Agent Framework documentation.

Semantic Kernel 1.x remains supported for critical bugs and security issues and is still the practical choice for Java (JDK 17+) and for systems built on its plugins, prompt templates, and vector store connectors. Microsoft Agent Framework reached general availability in April 2026 and receives most new agent features.

Agent Framework supports C# and Python in general availability, with a Go SDK in preview. Teams get strongly typed AI services that use the same dependency injection, telemetry, and patterns as the rest of the backend.

How Microsoft Agent Framework and Semantic Kernel Work

Both frameworks sit between application code and AI models, turning natural language intents into typed function calls, memory lookups, and multi-step execution graphs.

Semantic Kernel centers on a dependency injection container called the Kernel. Every agent needs a Kernel, and methods exposed to the model live in plugin classes marked with [KernelFunction]. Stateful business processes historically required the separate Process Framework on runtimes such as Dapr or Orleans.

Microsoft Agent Framework removes that ceremony. In .NET it builds on Microsoft.Extensions.AI primitives such as IChatClient, so agents plug into the same service collection as an ASP.NET Core API. One ChatClientAgent works with any chat client, and plain C# methods or Python functions can be passed as tools without a Kernel. Its main building blocks are listed in the table below.

The two stacks interoperate. From version 1.38 of the Python semantic-kernel package, any KernelFunction, including prompt templates, native methods, and vector search functions such as Azure AI Search, can become an Agent Framework tool with .as_agent_framework_tool().

Building blockWhat it doesNotes
AgentsLLM agents that call local code tools and MCP servers.ChatClientAgent in .NET, Agent in Python, plus Copilot Studio and A2A agents.
Harness AgentOpinionated agent for long-running, multi-step tasks.Planning, todo tracking, context compaction, file access, and tool approvals.
WorkflowsTyped, graph-based execution along explicit edges.Checkpointing, resumption, and human-in-the-loop pauses.
Integrations & middlewareOpenAPI, MCP, and Agent-to-Agent protocols.Middleware intercepts agent actions; Semantic Kernel uses filters.
Engineer's Take

Microsoft Agent Framework and Semantic Kernel are the natural choice when the backend runs on .NET or when Python services need strict typing, Azure integration, and consistent SDK conventions. Existing C# methods or Python functions become tools, and execution is traced through standard OpenTelemetry. Production failures usually come from identity, telemetry volume, and configuration rather than prompts: DefaultAzureCredential left in production causes credential probing and latency spikes, unsampled telemetry inflates storage bills, and Python Agent Framework does not load .env files automatically. For new projects we start on Agent Framework 1.0 and bridge existing Semantic Kernel plugins.

Diatom Enterprises favicon
Diatom engineering team
Microsoft Agent Framework at a Glance
Type

Open-source AI agent and orchestration SDKs

Languages

C# and Python; Java in Semantic Kernel; Go in preview

License

MIT

Strengths

Typed tools, graph workflows, Azure and OpenTelemetry integration

Compared with
Need a reliable tech partner?
20+

Years in business

80+

Senior engineers

200+

Projects delivered

What Microsoft Agent Framework and Semantic Kernel Are Used For

Engineering teams use Semantic Kernel and Microsoft Agent Framework to build tool-calling assistants, RAG pipelines, and stateful multi-agent workflows inside enterprise .NET, Python, and Java backends.

Grounded Enterprise Search

Vector store integrations such as Azure AI Search, exposed as callable search functions, let agents retrieve verifiable company records before answering staff or customers in internal portals.

Back-Office Automation

The Harness Agent handles long-running tasks that need planning, todo tracking, and file work, such as multi-document audits, with context compaction and remembered tool approvals so operators are not asked about repeated safe actions.

Human-in-the-Loop Workflows

Finance, healthcare, and logistics backends wire specialized agents along explicit workflow edges, checkpoint state, pause for a manager’s sign-off, and resume later, often with Azure Service Bus decoupling long tasks from HTTP timeouts.

Cross-Platform Agents and Java Modernization

MCP, A2A, and OpenAPI support let a C# or Python backend call external tools or Copilot Studio agents, while Java systems use Semantic Kernel’s native SDK instead of a separate Python service.

Semantic Kernel vs LangChain and LangGraph

When comparing Semantic Kernel vs LangChain and LangGraph, the LangChain stack favors Python-first prototyping and community breadth, while Semantic Kernel and Microsoft Agent Framework favor strict typing, .NET and Python parity, and Microsoft-maintained connectors. Agent Framework also succeeds AutoGen.

Criteriontech logo microsoftAgent Framework & Semantic Kerneltech logo langchainLangChaintech logo langchainLangGraph
LanguagesC# and Python (GA), Go preview; Java in Semantic Kernel.Python and JavaScript/TypeScript.Python and JavaScript/TypeScript.
Orchestration modelTyped graph workflows, Harness Agent, and Kernel plugins.Agent harness and composable building blocks.Directed state graphs with routing.
State & checkpointingNative workflow checkpointing, resumption, and human-in-the-loop.Inherits persistence from the LangGraph runtime.Built-in checkpointers and replay.
ConnectorsMicrosoft-maintained connectors with consistent typing.Very large community ecosystem with varying maintenance.Uses the LangChain integration ecosystem.
ObservabilityOpenTelemetry, Azure Monitor, and Application Insights.LangSmith and callback handlers.LangSmith tracing and deployment.
ProtocolsNative MCP, A2A, and OpenAPI support.Broad third-party tool wrappers and MCP adapters.Graph nodes can wrap external protocols.
Swipe the table to compare →

When to Choose Microsoft Agent Framework

Choose Microsoft Agent Framework and Semantic Kernel when your team writes C#, Java, or enterprise Python and deploys on Azure. Building on Microsoft.Extensions.AI means AI services follow the same dependency injection, middleware, and OpenTelemetry patterns as the rest of the backend, and typed graph workflows replace the less predictable conversational teams of AutoGen.

When Microsoft Agent Framework Is Not the Right Fit

A Python-only team prototyping against niche third-party APIs or experimental vector databases will often find a ready LangChain integration, though community packages vary in upkeep. Teams that want a standalone commercial tracing platform instead of configuring Azure Monitor often pair LangGraph with LangSmith. Production agents in Go should wait until the Go SDK leaves preview.

How We Approach Microsoft Agent Framework Projects

We build agent systems in .NET and Python with strict identity boundaries, deterministic workflow edges, and sampled OpenTelemetry observability from day one. Existing Semantic Kernel plugins are bridged into Agent Framework rather than rewritten, as part of our AI agent development services.

Explicit Cloud Identity

We replace DefaultAzureCredential in production with explicit credentials such as ManagedIdentityCredential, removing credential probing, latency spikes, and fallback security risks.

Telemetry Sampling and Clean Startup

Sampling rates are set before launch so Application Insights and Azure Monitor costs stay predictable, and Python services load environment variables explicitly because Agent Framework does not read .env files on its own.

Middleware Guardrails

Whenever agents call external MCP servers, third-party A2A agents, or non-Azure models, Agent Framework middleware or Semantic Kernel filters enforce data boundaries alongside content filters and system prompt rules.

Sound like the right partner?

Tell us about your project. We'll get back to you within 1 business day.

Frequently Asked Questions

Direct answers to common architectural and lifecycle questions when evaluating Semantic Kernel and Microsoft Agent Framework.

Did Microsoft Agent Framework replace Semantic Kernel and AutoGen?

Yes. Microsoft Agent Framework is the unified open-source successor to both, built by the same Microsoft teams. It reached general availability in April 2026, combining Semantic Kernel’s telemetry, typing, and connectors with multi-agent workflow orchestration. Semantic Kernel 1.x remains maintained, but new agent capabilities are developed in Agent Framework.

Should I use Semantic Kernel or Microsoft Agent Framework for new .NET projects?

Use Microsoft Agent Framework for new .NET projects that need agents or multi-step workflows. It replaces the Microsoft.SemanticKernel.Agents namespaces with Microsoft.Agents.AI and builds on standard Microsoft.Extensions.AI interfaces such as IChatClient. Semantic Kernel still makes sense when maintaining an existing 1.x application.

How long will Microsoft support Semantic Kernel?

Microsoft treats Semantic Kernel 1.x as a mature release and continues to fix critical bugs and security issues. Most new agent features go into Agent Framework, but Semantic Kernel remains relevant for existing enterprise applications, Java backends, prompt templating, and vector store integrations.

Can I reuse Semantic Kernel plugins and vector stores in Agent Framework?

Yes. From version 1.38 of the Python semantic-kernel package, any KernelFunction, whether built from a prompt template, a native method, or a vector search collection such as Azure AI Search, can be converted into an Agent Framework tool with .as_agent_framework_tool().

Which programming languages do they support?

Semantic Kernel supports C#, Python, and Java (JDK 17+). Microsoft Agent Framework supports C# and Python in general availability and offers a Go SDK in preview, which does not yet include declarative agents, CodeAct, or functional workflows. Java teams should keep using Semantic Kernel. For a wider comparison, see our guide to multi-agent frameworks.

The People Behind

A team of skilled engineers, architects, and delivery managers, dedicated to making every engagement a success.

Let's discuss your project