AutoGen vs LangChain: archived project or maintained stack?

A lot of developers still search for this pair as if it were a fair fight between two live frameworks. It isn't, and that changes the whole decision. One of them has been archived by its owner. The other is part of a stack that's still shipping new products. I'll call them the archived framework and the maintained stack from here on, because that's the clearest way to keep the difference in view.

Advice before the detail

If you're starting something new, pick the maintained stack, with its orchestration layer and the LangSmith platform for tracing, evals and deployment. If you already have multi-agent code running on the archived framework, keep it running while you plan a move, and use it as a reference for how its event-driven design works. If you're on the Microsoft stack and want a Microsoft-backed successor, the archived framework's own repository points you to Microsoft Agent Framework.

An archived project against a maintained stack

The archived framework's GitHub repository carries a caution: the project is archived, it's no longer actively maintained by Microsoft, and it advises against production use. Its docs site still looks active, which is exactly why people get caught out. You can read a polished guide, follow the examples, and never notice the notice at the top of the repository.

The maintained stack is in a very different place. Its docs and site are current, there are active product launches, and its homepage advertises an events programme. For anyone building something a business will rely on, that gap matters more than any feature on either list. A framework that won't get fixes is a risk you carry for the life of your project.

Where Microsoft points new users

The repository tells new users to start with Microsoft Agent Framework, which it describes as the enterprise-ready successor, and it offers a migration guide from the old project. That's the honest route if your team wanted the archived framework because it came from Microsoft. In that case, the useful comparison is Microsoft Agent Framework against the maintained stack.

The maintained stack doesn't need a successor. Its docs describe LangChain, LangGraph and Deep Agents as layers of one stack, so you move between layers rather than between projects as your needs grow.

Layer by layer

Area	AutoGen	LangChain
Project status	Archived; no longer actively maintained by Microsoft; not advised for production	Current docs and site, active product launches and an events programme
What comes next	Microsoft Agent Framework, with a migration guide	No replacement; one stack of layers, including Deep Agents
Architecture	AgentChat for conversational agents, Core for event-driven multi-agent runtimes, Extensions for external services	A create_agent harness of model, tools, prompt and middleware, with a graph layer for low-level orchestration
No-code	Studio, a web UI for prototyping agents on AgentChat	Fleet, a no-code agents product for building in natural language
Languages	Python 3.10+ and a .NET docs track	Python docs shown; frameworks presented as model-agnostic
Beyond the code	Extensions for MCP servers, the OpenAI Assistants API, Docker code execution and gRPC distributed agents; no hosted platform shown	LangSmith for observability, evaluation, deployment, sandboxes and an LLM gateway

How the code is organised

The archived framework is split into layered packages. AgentChat handles conversational agents. Core is an event-driven runtime for multi-agent systems, which the vendor says is built for scale. Extensions connect to outside services. It's a clean design, and it's part of why researchers liked it for studying how agents collaborate.

The maintained stack starts with a create_agent harness: you give it a model, tools, a prompt and middleware. When you need more control, LangGraph is the low-level option for mixing fixed, deterministic steps with agentic ones. The vendor's pitch is that you keep model choice and own your harness, so you aren't tied to a single model provider.

On languages, the archived framework supports Python 3.10 and up and has a .NET docs track. The maintained stack's docs are shown with a Python language selector, and its open source frameworks are presented as model-agnostic.

Building agents without writing code

Both sides have a no-code option, and they're aimed at different people. The archived framework's Studio is a web-based UI for prototyping agents without code, built on AgentChat. It's best thought of as a developer's sketchpad. The maintained stack lists Fleet, a no-code agents product designed so whole teams can build agents in natural language.

If your goal is to let non-developers in your business build their own agents, that difference counts. A prototyping tool on top of an archived project isn't something you'd roll out to a whole team.

What you get past the framework itself

The archived framework's Extensions cover MCP servers, the OpenAI Assistants API, Docker code execution and gRPC distributed agents. Its pages don't show a hosted platform to run or watch your agents.

The maintained stack adds LangSmith as a separate platform for observability, evaluation, deployment, sandboxes and an LLM gateway. The vendor describes the whole thing as one lifecycle: build, test, deploy and monitor. For a team shipping agents to customers, tracing and evals aren't extras, they're how you find out what went wrong. If you'd like to see more options in this space, our list of the best AI agent frameworks is a good next stop.

If you're on the archived framework today

• Note which parts of your code use AgentChat, Core and Extensions, since each maps differently to a new framework.
• Read the migration guide to Microsoft Agent Framework before you decide anything else.
• Compare that route with rebuilding on the maintained stack's harness and orchestration layer.
• Keep production changes small until the move is planned, because the old project won't get fixes.

Four more frameworks on the shortlist

Not sold on either? These four are the closest alternatives, and our page on other frameworks to use instead of LangChain goes into more.

CrewAI

Best for: Business and engineering teams that want governed, centrally managed agents, on an enterprise platform with an open source edition.

LlamaIndex

Best for: Teams building agents over complex documents such as invoices and claims, with LlamaParse for parsing, extraction and indexing.

LangGraph

Best for: Developers who need fine-grained control over stateful, long-running agent logic, including memory and human-in-the-loop steps.

Semantic Kernel

Best for: Enterprise developers on the Microsoft stack in C#, Java or Python who want to add AI models to existing codebases.

Do you make an AI agent framework or agent platform? Tool makers can ask to be considered for a featured place on this page.

Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.

Related: the python page.

Frequently asked questions

Is AutoGen still maintained?

Its GitHub repository says the project is archived and no longer actively maintained by Microsoft, and it advises against production use.

What replaces AutoGen?

The repository points new users to Microsoft Agent Framework, described as its enterprise-ready successor, and offers a migration guide.

Is LangChain the same as LangGraph?

They're layers of one stack. The create_agent harness sits on top, and the graph layer is the low-level option for orchestrating deterministic and agentic steps.

Can I build agents without code on either?

The archived framework has Studio, a web UI for prototyping. The maintained stack lists Fleet, a no-code agents product for whole teams.

Should I still learn the archived framework?

It's worth reading for its event-driven multi-agent design, but don't start a new production project on it.