Microsoft Agent Framework vs Semantic Kernel (2026)

For new agent code, use Microsoft Agent Framework (MAF). Microsoft built it as Semantic Kernel's successor, with the same team, and the Semantic Kernel README now opens by saying "Semantic Kernel is now Microsoft Agent Framework." Semantic Kernel is still maintained, so existing production code does not have to move this quarter. But new agent features go to MAF, and the migration is mostly renames. Sources checked 2026-11-09.

Quick comparison

How they differ

Who gets new features

Microsoft's October 2025 FAQ told developers to think of Agent Framework as "Semantic Kernel v2.0," built by the same team. Semantic Kernel 1.x keeps getting critical bug and security fixes, and a few existing features were to reach GA, "but the majority of new features will be built for Microsoft Agent Framework." The support promise was at least one year after MAF became generally available. MAF 1.0.0 landed on PyPI on April 2, 2026, so that promise runs to at least April 2027.

Semantic Kernel has not stopped releasing. Python 1.47.0 shipped on November 5, 2026, dropping Python 3.10 and fixing MCP prompt arguments, history truncation and several vector store connectors. .NET 1.82.0 followed on November 7. Those are mostly fixes and dependency updates, though, not new agent features.

Less wiring in Agent Framework

The migration guide shows where the code shrinks:

• No Kernel. In Semantic Kernel every agent depends on a Kernel instance. In MAF you pass a chat client to Agent(...) or call client.as_agent(instructions=...).
• One agent type. MAF's Agent covers Foundry, OpenAI Chat Completions and OpenAI Responses. There is no separate class per service.
• Tools are functions. A Python function's name and docstring become the tool name and description. You don't need decorators or a plugin class.
• Sessions come from the agent. agent.create_session() replaces creating ChatHistoryAgentThread or a service-specific thread yourself.
• invoke becomes run. A non-streaming call returns one AgentResponse with .text and .messages. Streaming is run(..., stream=True).
• Typed options. Settings move from KernelArguments and execution-settings objects to a default_options / options dict with provider-specific TypedDicts.

What you keep from Semantic Kernel

You do not have to rewrite everything at once. From semantic-kernel 1.38, as_agent_framework_tool turns a KernelFunction into an MAF tool. That works for prompt-template functions, @kernel_function methods, and the create_search_function of any Semantic Kernel vector store collection. So the usual path is to move the agent loop first and bring your plugins and vector search along as tools.

The cost of being current

MAF moves fast. The Python side shipped 1.21.0 through 1.25.0 between October 8 and November 7, 2026, and 1.25.0 marked seven changes as breaking: Python 3.11 minimum, bounded MCP discovery and 30-second timeouts, workflow agents bound to one authorization scope, renamed workflow spans and more. Each came with a migration note. Pin versions and read release notes before you upgrade. Semantic Kernel changes less, which is the remaining argument for leaving a stable SK service alone for now.

When Semantic Kernel is still the right call

• You run it in production and nothing on your roadmap needs MAF-only features. Plan the migration, but you don't have to rush it.
• You need Java. MAF has no Java SDK. Its third language is Go.
• Your app relies on Semantic Kernel's Process Framework or a connector you haven't checked in MAF yet. Confirm MAF has an equivalent before you switch.

For other framework pairings see LangChain vs LangGraph and LangGraph vs Google ADK, or browse the frameworks category, which includes AG2, the AutoGen offshoot.