LlamaIndex vs Semantic Kernel: what changed with Agent Framework

If you are starting a new agent project on Microsoft's stack, the comparison has moved. Microsoft now describes Microsoft Agent Framework as the successor to Semantic Kernel, and version 1.0 shipped for .NET and Python in April 2026. For most new work, the real choice is between LlamaIndex and Agent Framework, with Semantic Kernel as the option for code you already run.

The second factor is language. LlamaIndex is now effectively a Python choice, while Semantic Kernel is the only one of these three projects with a Java SDK. Your team's language can settle the question before any feature comparison does.

This comparison is based on each project's documentation and repositories as of September 24, 2026. It is not a benchmark or a hands-on test.

Where Semantic Kernel stands now

The Semantic Kernel repository now opens by calling Microsoft Agent Framework its "enterprise-ready successor" and points readers to a migration guide. Microsoft's announcement of Agent Framework 1.0 describes stable APIs and a commitment to long-term support.

Semantic Kernel is not abandoned. In an October 2025 post, Microsoft said it would keep fixing critical bugs and security issues, while most new features would go to Agent Framework. It committed to supporting Semantic Kernel for at least one year after Agent Framework became generally available. Releases are still shipping: .NET 1.80.1 appeared on September 3, 2026.

That changes how to read older comparisons, including the earlier version of this article. The "Agent Framework 1.0 by Q1 2025" milestone it described belonged to Semantic Kernel's own agent package, which Microsoft declared generally available for .NET and Python in April 2025. That package is a different product from today's Microsoft Agent Framework. A comparison that covers Semantic Kernel's agent features without mentioning the successor describes a line Microsoft is no longer extending.

Choose by language and job

Language support is narrower than older comparisons suggest

LlamaIndex's framework overview still says it is available in Python and TypeScript. The LlamaIndex.TS repository, however, carries a notice that the project is deprecated and no longer maintained, and it was archived on April 30, 2026. The Python llama_index repository is active and MIT-licensed, with release 0.14.25 published on September 21, 2026. For a new project, treat LlamaIndex as a Python dependency.

Semantic Kernel supports C#, Python and Java, but not with equal depth. Microsoft's supported languages table shows Java missing OpenTelemetry logs and several model connectors, including Anthropic, Amazon Bedrock and Ollama. Check that table against the services you plan to call.

Agent Framework's overview covers .NET and Python, with a Go version in public preview. It does not list a Java version. A Java team that wants Microsoft's agent tooling faces a real tradeoff: Semantic Kernel for Java works today, but it sits on the maintenance track.

LlamaIndex is built around your documents

LlamaIndex's strength is getting your own data into a form a model can use. Data connectors, which it calls readers, load sources into a common document format. Each one is a separate package named llama-index-readers-<source>, so you install only what you need. LlamaIndex data connectors

It is no longer only a retrieval library, though. The framework overview also covers agents and multi-step workflows. The old claim that LlamaIndex handles data while Semantic Kernel handles agents is too tidy to guide a decision now.

Keep the open-source framework separate from the company's paid products. The LlamaIndex homepage now leads with LlamaParse, a hosted document-parsing service. As of September 24, 2026, its pricing page lists a free plan with 10,000 credits a month and paid plans priced per month, with credits valued at $1.25 per 1,000. Parsing cost depends on the tier: LlamaParse's credit documentation lists 1 credit per page for its Fast tier, 3 for Cost-effective, 10 for Agentic and 45 for Agentic Plus. You can use the MIT-licensed framework without it. If your documents are complex PDFs with tables, though, test whether you would need the paid parser, because that changes your running cost.

Semantic Kernel and Agent Framework are built around actions

Semantic Kernel's design starts from functions. You describe existing code as plugins, and the model calls them to act on a request. For retrieval, it offers vector store connectors and a text-search layer for RAG. Microsoft still marks the vector store feature RC in Python and preview in Java. Semantic Kernel vector stores

Agent Framework keeps that orientation and simplifies it. Its overview describes agents that call tools and MCP servers, graph-based workflows with explicit execution paths, and support for providers including Azure OpenAI, OpenAI, Anthropic and Ollama. Microsoft also places responsibility for safety mitigations, testing and data boundaries on the developer. Using a Microsoft framework does not make an application compliant by itself.

Both can sit in one Python application

You don't have to pick one framework for everything. In Python, Agent Framework accepts a plain function as a tool, according to its migration guide. That means you could wrap a LlamaIndex query in a function and hand it to an Agent Framework agent. This combination is not something we have built or tested. It follows from the documented interfaces, so prototype it before relying on it.

Two frameworks mean two sets of dependencies and upgrades. Combine them only when one of them clearly falls short on its own.

If you already run Semantic Kernel

There is no emergency. The migration guide shows what changes: .NET namespaces move to Microsoft.Agents.AI, agents no longer need a Kernel instance, functions register directly as tools, and Invoke calls become Run. In Python, existing KernelFunction objects can be converted into Agent Framework tools with Semantic Kernel 1.38 or later, which allows a gradual move.

Plan the migration around the features you need. If a capability you want only arrives in Agent Framework, that is your trigger to move. If your application is stable and receives security fixes, the support commitment gives you time to schedule the work.

Where LangChain fits

Many searches for this comparison add LangChain. It is the most practical of the three names for TypeScript teams, since it maintains both Python and JavaScript libraries under the MIT license. Its agents are built on LangGraph, which provides persistence and human-in-the-loop steps.

We have not listed GitHub stars. They measure attention rather than fit, and the Semantic Kernel repository now points visitors to a different project. A dated star count can't tell you whether a framework supports your language, data sources and deployment.

Test with your own documents and one real task

Before committing, build the same small slice with each candidate. Use a real set of your documents and one task an agent must complete, such as answering a question and then filing the result somewhere. Record:

• How much code you needed to load and query the documents.
• Whether the answer cited the right source passages.
• How the framework reports a failed tool call.
• What you would pay for parsing, models and hosting at your expected volume.
• Which parts you would need to rewrite if you switched later.

Pick the framework that handles your hardest part with the least custom code, in a language your team already maintains.