I Checked 20 Agent Frameworks for a Published Support Window. Zero Have One.

TL;DR: Probably, but nobody has promised you anything in writing. I pulled release history, repo-root policy files and governance for 20 agent frameworks on 15 August 2026: zero publish a version-support window or an end-of-life date, 19 of 20 are governed by a single vendor, and 17 of the 19 with usable release histories shipped something in the last 90 days. The one that did not is AutoGen, which has 60,442 stars and has not cut a release since 30 September 2025.

AutoGen is the first major agent framework to be sunset by the company that built it, and it did not go quietly into obscurity. Its GitHub repository still carries 60,442 stars and 444 contributors by GitHub's own contributors API, plus 992 open issues. Its README now opens with a banner: "AutoGen is now in maintenance mode. It will not receive new features or enhancements and is community managed going forward."

Its successor, Microsoft Agent Framework, published its first beta to PyPI on 1 October 2025 and its 1.0.0 stable release on 2 April 2026. Microsoft's own documentation calls it "the direct successor" to both AutoGen and Semantic Kernel, "created by the same teams."

Maintenance mode is not death. AutoGen still takes bug fixes and security patches, and the fork line is very much alive: AG2 shipped v1.0.2 on the day I ran this, and its README says the project is "maintained by a dynamic group of volunteers from several organizations," naming Chi Wang and Qingyun Wu as administrators. What changed is the roadmap, not the pulse. Here is how to tell the difference before it matters to you.

How do you check whether your framework is next?

1. Find the last feature release, not the last release. Pull the version list from PyPI or npm and find the most recent major or minor bump, not the most recent publish. Failure mode: a project in maintenance mode still ships patches, so "last commit" and "last release" both look healthy while the roadmap is already closed. AutoGen's last publish was 0.7.5 on 30 September 2025. Its last minor bump was 0.7.0, two months earlier. A dashboard watching commits would have shown green through the whole sunset.

2. Read the SUPPORT.md instead of checking that it exists. Failure mode: treating file presence as a policy. Four of the 20 repos carry a support, migration or maintainers file at the root beyond the boilerplate SECURITY.md, and none of them states a supported-version window. AutoGen's SUPPORT.md reads, in full, under the heading "Microsoft Support Policy": "Support for this project is limited to the resources listed above." The resources listed above are GitHub Issues and GitHub Discussions.

3. Check whether support is conditional on buying something else. Failure mode: assuming an MIT license implies open-source support from the vendor. Microsoft Agent Framework's SUPPORT.md says its AI support team handles issues "for customers under a Unified support agreement when the issue arises from usage of Azure AI services," and that otherwise the SDK "is treated as a purely open-source tool" that "Microsoft's support organization will not handle." The successor framework has more support than AutoGen. It also has a price of entry.

4. Ask who governs it, then ask what they have already retired. Failure mode: counting stars instead of counting owners. Nineteen of the 20 repos are owned by a single commercial vendor. One, AG2, describes volunteer governance. None is under a software foundation. Three of the 20 belong to Microsoft (AutoGen, Semantic Kernel, Agent Framework), and Microsoft has already put one of the three into maintenance mode and folded a second into the third.

5. Look for a migration path before you need one. Failure mode: discovering a project has never performed a migration at the moment it has to. Four in my sample publish a documented migration from a predecessor that I fetched and read: AutoGen to Agent Framework, Semantic Kernel to Agent Framework, Haystack 1.x to 2.x via its root MIGRATION.md, and LangChain's versioned upgrade docs. A vendor that has moved users across a breaking boundary once has proven it can. The rest are unproven, which is not the same as bad.

What did the 20-framework sweep show?

I recorded, for each of AutoGen, AG2, Microsoft Agent Framework, Semantic Kernel, LangChain, LangGraph, OpenAI Agents SDK, CrewAI, LlamaIndex, Google ADK, AWS Strands Agents, Claude Agent SDK, Pydantic AI, smolagents, Agno, LiveKit Agents, Haystack, Letta, Mastra and the Vercel AI SDK: last release date, last major or minor release date, repo-root policy files, owner, and whether any support window is published.

Nineteen had usable release histories. I dropped the Vercel AI SDK from the recency count because the ai package name on npm has a discontinuous history predating Vercel's ownership, with a first publish in 2014, which makes version ordering meaningless without hand-reconstruction.

Of those 19, 17 shipped a release in the last 90 days and 13 shipped a major or minor bump in the same window. The two that shipped nothing are AutoGen (318 days since its last publish) and Letta (93 days). LangChain and LangGraph both shipped patches three days before I ran this but last bumped a minor version on 12 May 2026, 94 days back. That is a healthy project on a slow feature cycle, not a dying one, and the distinction is the whole point of step one.

The support-window count is the finding: zero of 20. Not one publishes a date after which a version stops receiving fixes. Compare that with the databases and language runtimes these same teams depend on, most of which publish end-of-life calendars years out.

Install counts would have been the obvious thing to add here and I left them out. PyPI and npm download numbers are inflated by CI runners and mirrors in ways that vary wildly by project, so a ranking built on them measures build pipelines more than adoption.

What does this not prove?

It does not prove any of these projects are at risk. Nineteen of twenty are shipping, most of them weekly, and a single-vendor project with a full-time team is often better maintained than a foundation project with none. Absence of a published support window is a documentation gap, not a health signal. It just means the thing you would check is not checkable.

It also does not prove AutoGen users were stranded. They got a successor, a migration guide, and a community fork with an active v1.0. That is a better outcome than most sunsets. The lesson is about timing: the announcement and the last feature release both landed while most surveyed teams already had agents in production. LangChain's report puts that at 57.3% of 1,340 respondents. Two caveats, both real. LangChain is a framework vendor surveying framework users, which is the same structural problem as vendors running most of the AI coding productivity studies. And the survey was fielded 18 November to 2 December 2025, then published on 12 June 2026, so a headline that reads as mid-2026 describes late 2025. In this field that is a long time.

What metric should you track?

One number per framework: days since the last major or minor release. Not days since last commit, which stays green through a sunset. Not stars, which are a lifetime accumulator and went up while AutoGen was being retired. Pull it from the PyPI or npm JSON API, which takes one request per package and no key, and put it on the same quarterly review as your dependency audit.

Set the line at 180 days and treat a crossing as a prompt to read the repo, not as an alarm, and read the LICENSE while you are in there, because 4 of 35 repos in the "best open source AI agent" roundups are not under an OSI-approved license at all. In my sample that line flags three of nineteen: AutoGen at 382 days since a feature release, Letta at 242, and LlamaIndex at 340, the last of which reflects a 0.x versioning convention where features ship inside patch numbers rather than a stalled project. The metric is a tripwire, not a verdict. Which is roughly the same posture the right amount of leash requires everywhere else in this stack, as the gap between agent demos and real work keeps demonstrating. And when the framework does move under you, it is worth knowing in advance that the contract you signed almost certainly does not cover the consequences.