Skip to content

AI Agent Frameworks Compared: A GitHub Health Check on LangChain, CrewAI, AG2 and the AutoGen Shutdown

· El Mehdi

“Best AI agent framework” roundups get written once, ranked by GitHub stars, and never revisited — which is a problem in a category where the single most-starred project in our sample just put a red “archived” badge on its own README. We pulled live data straight from GitHub for seven of the most-recommended AI agent frameworks — LangChain, LlamaIndex, CrewAI, smolagents, AG2, Microsoft’s original AutoGen, and AutoGen’s official successor, Microsoft Agent Framework — on November 9, 2026, to see which ones are actually shipping right now.

What we checked, and how

For each repository we pulled star count, fork count, primary license, and the most recent tagged release directly from GitHub. GitHub’s repository API bundles open issues and open pull requests into a single “open issues” number, so to split them we separately queried each repo’s open-issues-only count and subtracted it from the combined total to get open PRs. Every figure below reflects what GitHub itself returned on November 9, 2026, not an estimate.

The headline: the framework with the most stars has a self-declared “archived” banner

Of the seven projects we checked, Microsoft’s original AutoGen repository has the most stars — 61.8k, more than LangChain’s own agent-focused repos, CrewAI, or AutoGen’s direct successor. It is also the only one of the seven with no tagged release in over a year. Its README now opens with a status badge reading “Archived” and this notice:

“⚠️ Project archived and no longer maintained. This project has been archived and is no longer actively maintained by Microsoft. Microsoft is not providing bug fixes, security fixes, new features, or support for this project. […] New users should start with Microsoft Agent Framework.”

One nuance worth being precise about: this is a README-level notice, not a platform-level one. GitHub’s own repository metadata still shows archived: false for microsoft/autogen — the issue tracker and pull requests are technically still open, and the repo was pushed to as recently as October 28, 2026. What’s actually true is narrower than “frozen”: Microsoft has stopped tagging releases (the last one, python-v0.7.5, shipped September 30, 2025 — over 13 months ago) and has redirected new and existing users to a different, actively-developed repository.

That successor, Microsoft Agent Framework, tells a very different story in the numbers: a tagged release two days before we pulled this data (python-1.25.0, November 7, 2026), a push the same day we checked, and a healthier open-issue-to-PR ratio than the project it replaces. It’s also, by star count, the smallest of the seven we checked — 14.6k stars versus AutoGen’s 61.8k — a reminder that a star count is a record of past attention, not a measure of what’s maintained today.

GitHub health, side by side

FrameworkStarsForksOpen issuesOpen PRsLicenseLatest releaseStatus
LangChain149.3k25.1k593137MITlangchain-core==1.7.6 (Nov 5, 2026)Active, daily release cadence
CrewAI60.7k9.1k492721MIT1.15.39 (Nov 9, 2026)Active, daily release cadence
Microsoft AutoGen (original)61.8k9.4k572600MIT (code) / CC-BY-4.0 (docs)python-v0.7.5 (Sep 30, 2025)Archived per README; superseded
Microsoft Agent Framework (successor)14.6k2.6k672233MITpython-1.25.0 (Nov 7, 2026)Active, official AutoGen successor
LlamaIndex52.9k8.5k389942MITv0.14.25 (Sep 21, 2026)Active, ~7 weeks since last tag
smolagents (Hugging Face)30.3k3.2k417636Apache-2.0v1.26.0 (May 29, 2026)Active, slowing (~5.4 months since last tag)
AG25.1k0.8k2465Apache-2.0v1.3.0 (Nov 4, 2026)Active, small team

Framework-by-framework notes

LangChain — the biggest, and still shipping daily

LangChain’s repository has 1,419 tags, because it’s a monorepo that publishes dozens of interdependent packages (core, partner integrations, text splitters, and more) on their own schedules — “latest release” here means a sub-package shipped hours before we checked, not a single project-wide version number the way AG2 or CrewAI use one. With 149.3k stars and a push timestamped the same day we pulled this data, it’s comfortably the most active project in this comparison, and its issue-to-PR ratio (593 open issues against 137 open PRs) is the most triaged-looking of the seven.

CrewAI — frequent point releases, a heavier PR backlog

CrewAI tagged version 1.15.39 the same day we checked, continuing a pattern of frequent small point releases. Its 721 open PRs against 492 open issues is a backlog worth watching, though it’s the kind of ratio you’d expect from a fast-growing project pulling in more community contributions than its maintainers can merge — not, by itself, a sign of abandonment the way it was for a project that had actually stopped tagging releases.

LlamaIndex — the widest PR backlog we measured

LlamaIndex has the most lopsided ratio of the seven: 942 open pull requests against 389 open issues, meaning open PRs outnumber open issues by better than 2 to 1. The project is still active — pushed to on November 8, 2026 — but its most recent tagged release, v0.14.25, is about seven weeks old as of this writing, longer than LangChain’s or CrewAI’s gap.

smolagents — Hugging Face’s framework, cooling off since spring

smolagents’ last tagged release, v1.26.0, shipped May 29, 2026 — roughly five and a half months before we checked, the longest gap among the actively-maintained frameworks in this table (AutoGen’s gap is longer, but AutoGen is no longer being actively maintained at all). Commits are still landing — the repo was pushed to on November 3, 2026 — but if you’re tracking this one, the release cadence has clearly slowed from wherever it started.

AG2 — small, independently governed, still shipping

AG2 traces back to late 2024, when a group of AutoGen’s original contributors split off from Microsoft and forked the AutoGen codebase under independent governance, eventually landing on the ag2ai/ag2 repository and the ag2, autogen, and pyautogen PyPI package names. (Microsoft engineers have publicly disputed how that split is characterized, including whether to call it a departure of “original creators” versus “part of the team” — the repositories and release history are what we can verify directly, and we’re reporting those rather than relitigating the dispute.) At 5.1k stars it’s by far the smallest project here, but it tagged v1.3.0 on November 4, 2026, five days before we checked, and its open issue count (24) is in line with its smaller scale.

Microsoft AutoGen — see above. The README says it’s done.

Covered in detail above: a self-declared archived status, a final tagged release from September 2025, and an explicit redirect to Microsoft Agent Framework for new and existing users alike. Worth repeating: the GitHub platform itself hasn’t flipped the technical “archived” switch (issues and PRs can still be opened), but Microsoft’s own documentation is unambiguous about where ongoing development has gone.

Microsoft Agent Framework — the successor, and it’s the one actually shipping

Created in April 2025 and positioned as the unified, enterprise-supported successor to both AutoGen and Semantic Kernel, Microsoft Agent Framework reached general availability earlier in 2026. Of the seven repositories we checked, it has the most recent tagged release relative to our check date — two days prior — and the most favorable issue-to-PR ratio of the group apart from LangChain’s. If you’re evaluating this category today, it’s the Microsoft-maintained option that is actually under active, production-track development; the original AutoGen repository is not.

How to actually use this before picking a framework

Star count is a record of attention, not maintenance — the most-starred project in our sample is the one whose own maintainers say to use something else. What’s more informative: the date of the most recent tagged release relative to today, not a README’s “actively maintained” badge written at some earlier point and never revisited; whether open PRs dwarf open issues, which can mean an unreviewed contribution backlog but can also just mean a popular, growing project outpacing its maintainer bandwidth — so read it alongside release cadence, not in isolation; and, for anything built on Microsoft’s agent tooling specifically, whether you’re pointed at the original AutoGen repository (no longer actively developed, per its own README) or Microsoft Agent Framework (its actively-developed successor) — the two are easy to confuse by name alone.

What we left out, on purpose

We did not report total commit counts or contributor counts, because we could not pull them through the same consistent method across all seven repositories — we’d rather leave a column out than mix figures gathered different ways. We also did not independently verify Microsoft Agent Framework’s exact general-availability date beyond what’s stated in its own repository and documentation; “earlier in 2026” is deliberately approximate rather than a date we couldn’t confirm firsthand. Star, fork, issue, and PR counts move constantly; treat every number above as a snapshot of November 9, 2026, not a running total.

Leave a Reply

Your email address will not be published. Required fields are marked *