Which AI Agent Frameworks Are Still Maintained? We Checked 9 GitHub Repos (October 2026)

Lists of “best AI agent frameworks” rarely tell you whether anyone is still working on the code. I think that matters more than any feature list: a framework that stopped receiving fixes in the spring won’t get security patches in the autumn, and agents tend to hold API keys and access to other systems. So instead of repeating vendor claims, I went to the source code and counted.

Live tracker: the numbers in this article are a snapshot from mid-October 2026. For the current status of these frameworks, and more than 100 open-source AI tools in total, each with a 0 to 100 health score, see the AI Tool Health Index, which refreshes from GitHub automatically. Two of these are also compared side by side: CrewAI vs LangGraph and n8n vs Dify. The tracker counts every commit, including merge commits, and uses a 91-day window, so its figures run a little higher than the ones below.

I haven’t built anything with these frameworks myself yet. I checked them the way you’d check any dependency before trusting it: by looking at who is still working on the code.

The short version: seven of the nine projects are healthy. Two are not, and both still show up on recommendation lists: Flowise, which has been archived, and Microsoft’s AutoGen, which Microsoft now describes as archived and no longer maintained. If you’re choosing today, I’d rule both out.

The numbers

* Counted on 15 October. The next day, 16 October 2026, Microsoft added one more commit to AutoGen: a README change declaring the project archived (see below).

Repositories measured: n8n-io/n8n, langgenius/dify, microsoft/agent-framework, langflow-ai/langflow, All-Hands-AI/OpenHands, crewAIInc/crewAI, langchain-ai/langgraph, FlowiseAI/Flowise and microsoft/autogen.

What stands out

Two projects have stopped

Flowise’s README now opens with the line “Flowise has been archived” and links to a “Future of Flowise” discussion. The last commit we found is a README update on 13 August 2026; the four commits before it, from 7 to 10 August, were bug and permission fixes. Activity had already thinned out before that: 104 commits in six months, with only four distinct author names in the last three months.

AutoGen has now gone the same way. When we first checked on 15 October, its README carried a “Maintenance Mode” warning and there had been no commits since 6 April 2026. On 16 October 2026 Microsoft replaced that banner: the README now says the project is “archived and no longer maintained”, that Microsoft is “not providing bug fixes, security fixes, new features, or support”, and that the repository stays available for reference. New users are pointed to Microsoft Agent Framework, with a migration guide for existing users. Meanwhile, Agent Framework logged almost a thousand commits in 90 days.

If you are starting a new project, neither Flowise nor AutoGen is a sensible base. If you already run one of them, the practical question is when to migrate, not whether. We cover safer starting points in our guide to free AI agents for beginners.

The visual builders are the busiest

n8n and Dify were far ahead of everything else. Part of that is size: both are large products with a front end, back end, integrations and documentation in a single repository, so a raw commit count flatters them compared with a focused library like LangGraph. Still, Dify having 229 distinct author names in three months means a lot of people are fixing and extending it.

Low commit counts are not automatically bad

LangGraph’s 173 commits look small next to n8n’s 3,751, but LangGraph is a library with a narrow job, and some LangChain work happens in other repositories. Use these numbers to spot projects that have stopped, not to rank healthy ones against each other.

Licences: “open source” doesn’t mean the same thing everywhere

We read the licence file at the root of each repository. Three of them come with conditions that matter if you plan to sell a product built on top:

This is a summary, not legal advice. Licence files can change between versions, so read the one that ships with the release you deploy.

A second check: how much do you run to self-host?

For the four visual builders, we also counted the services defined in each project’s own Docker Compose file, using the default settings. This gives a rough idea of how many moving parts a beginner takes on.

There is an obvious irony here. Flowise was the easiest to run, and it is the one that has stopped. Dify gives you the most out of the box, but you are also looking after sixteen containers. Langflow is the lightest of the actively maintained builders we checked. We counted the files; we did not time an installation, so we can’t say how long setup takes on a real server.

What this means if you’re choosing a tool

• Starting fresh with a visual builder: n8n, Dify and Langflow are all actively developed. Choose by what you need to build and how much infrastructure you’re willing to run, not by popularity.
• Writing code: LangGraph, CrewAI and Microsoft Agent Framework are all active. If you’re on AutoGen, Microsoft publishes a migration guide to Agent Framework.
• On Flowise today: plan a migration. An archived project will not receive security fixes, and agents often hold API keys and access to other systems. Our AI agent safety guide explains why that combination is risky.
• Selling a product: read n8n’s and Dify’s licences before you commit.

My take

If I were picking a framework for a new project today, I’d use the maintenance check as a filter first and the feature comparison second. That rules out Flowise and AutoGen straight away. Among the rest, I’d choose n8n or Dify for visual building depending on whether the job is automating apps or answering from documents, and LangGraph or Microsoft Agent Framework if the team writes code. Before committing to n8n or Dify for a commercial product, I’d read their licences with a lawyer, because neither is a plain permissive licence.

The bigger lesson is that this market moves fast enough for a project to go from recommended to archived within a year. It’s worth repeating the five-line check at the end of this article before you adopt anything, and again every few months.

Limits of this analysis

• Commits measure activity, not quality. Many small commits can be less meaningful than a few large ones.
• Author names are an approximation of contributors: one person may commit under two names, and our “bot” filter only removes names that contain the word.
• Repositories differ in scope. A monorepo counts work that a library spreads across several repositories.
• We measured the default branch only, on one day. Numbers will have moved by the time you read this.
• We did not install, run or benchmark any of these tools.

How to repeat this yourself

You only need git. For any repository:

git clone --filter=blob:none --no-checkout https://github.com/OWNER/REPO
cd REPO
git log -1 --format=%cs
git log --no-merges --since=90.days --oneline | wc -l
git log --since=90.days --format='%an' | grep -vi bot | sort -u | wc -l

Run it before you adopt a framework and again every few months. If the numbers suddenly drop to near zero, read the README and issue tracker for an announcement.

Related reading: AI agent infrastructure explained · AI coding agents in 2026 · AI tool pricing compared

More from EvalAITools: Does hybrid search help RAG? Our test · MiniCPM5-2B for local agents · Framework defaults for approval and step limits · The same check for 14 RAG tools · The same check for 12 local LLM tools