60,871 stars shipped November 12. The 61,809-star repo did not.

Install for the job. The star sort fails. At 03:46 UTC on November 14, 2026, crewAIInc/crewAI had 60,871 stars and a latest release, 1.15.41, from November 12. microsoft/autogen had 61,809 stars, 56 more, and its latest release was still python-v0.7.5 from September 30, 2025. That is 410 days. The README says the project is archived. The GitHub field archived was false.

LangGraph 1.2.22 published at 01:43 UTC the same morning, with 44,932 stars and 288 contributors, the second-lowest contributor count in this set of 11. The release notes are performance fixes and a streaming fix, not a new orchestration model. The checkpoint was already the reason to pick it.

This page is the install decision. The November 5 GitHub snapshot is stars and one tag. The November 13 protocol matrix is MCP and A2A. LangGraph vs CrewAI vs the OpenAI Agents SDK is three bills, read November 10. None of those is today's count, and none of them is the job.

Which framework do you install?

Match the failure you cannot afford. A crashed run, a crew of roles, a handoff you host yourself, a Google Cloud deploy, or a TypeScript workflow are five different purchases. Stars do not distinguish them.

A harness is the loop, the tools, the memory, and the stop rule around a model. These libraries are how you build that loop. They are not the model bill. The November 10 comparison prices LangSmith seats, CrewAI AMP, and the model API for three of these rows. This note does not reprice them.

What did GitHub return at 03:46 UTC?

Eleven repositories. Stars, forks, and the latest release are copied from the API. Contributor counts are the last page of the contributors list with anon=0, so anonymous commits are out. Calendar age counts UTC dates from the release date through November 14, 2026.

The contributor sort is not the star sort. LlamaIndex leads contributors at 476 and is third in stars. Pydantic AI is second in contributors, 475, and ninth in stars. LangGraph shipped this morning and has 288 contributors. AG2 has 432 contributors and 5,063 stars. A contributor count is accounts with commits the API listed. It is not the size of the team that will answer your issue. The November 11 issue-health note is the closer reading of who actually closes bugs.

Licenses and languages, same pass:

Mastra's SPDX field is NOASSERTION because the detector did not pick a single id. The root LICENSE.md, read from the license API at 03:47 UTC, puts Apache 2.0 on content outside the ee/ directories and points those directories at a separate license. That ee/ file was not opened. AutoGen's root LICENSE is the CC BY 4.0 text, which is why GitHub says CC-BY-4.0. This note did not re-check the PyPI classifier.

Microsoft Agent Framework is not a Python-only repository. C# is 38.5% of the bytes. AutoGen is the same shape, with a TypeScript slice on top. LlamaIndex's second language is notebooks. AG2's is MDX, which is docs in the tree, not a second runtime.

What can pause, and who will host it?

Durable state, a handoff, MCP, tracing, and a hosted runtime are five different wires. The November 13 matrix already did MCP client, MCP server, and A2A for a wider set. The cells below are only what was on a page fetched today. A dash means that page did not document the feature. A dash is not a no.

LangGraph's persistence page splits the two memories. A checkpointer saves the thread so you can continue a conversation, resume after an interrupt, or recover from a failure. A store keeps facts across threads. The overview says LangGraph is the orchestration runtime and LangSmith is where tracing, evaluation, and deployment live. Agent Server can run in LangChain's cloud, in your VPC with LangChain's control plane, fully self-hosted, or as containers with no control plane. The page also says you can deploy an ADK agent through that server. That is LangChain's claim about ADK, not Google's deploy page.

CrewAI's 1.15.41 flows page says @persist writes state so a flow can restart. The default it names is SQLiteFlowPersistence. The pricing page lists Basic as Free, with a visual editor, GitHub integration, and 50 workflow executions a month. Enterprise is Custom. The sentence under it says you can deploy on CrewAI cloud, your own VPC, or your own infrastructure. There is no dollar figure on that card. This note does not invent one.

The OpenAI Agents SDK handoff page is delegation to a specialist. The session page is conversation memory, including a resume after an approval pause, with SQLite, Redis, and SQLAlchemy backends. That is not LangGraph's checkpoint of every superstep. The tracing page says tracing is on by default, and off for organizations on a Zero Data Retention policy. The Agents API guide is a different product: hosted orchestration, with the SDK described as the thing your application runs. The Python package's latest GitHub release is still v0.24.0 from October 15, 30 days before this pull. Do not describe the SDK and the hosted API as one install.

Google's deploy page names four places: Agent Runtime on Google Cloud, Cloud Run, GKE, and a container you run anywhere, including a machine with no Google Cloud connection. The session page says the in-memory services are for local tests and lose everything on restart, and that database and cloud backends exist. The eval page is unusually concrete for this set. It scores the trajectory, the tool calls along the way, and the final response, with test files, eval sets, and a conformance command against a recorded baseline. The MCP page is both directions: McpToolset consumes a server, and to_mcp_server exposes ADK tools.

Microsoft's checkpoint page saves state at the end of each superstep so a long run can pause, resume, or move machines. The Python storages it names are in-memory, a local directory, and Azure Cosmos DB. Handoff sits in the list of built-in orchestration names, next to Concurrent, Sequential, GroupChat, and Magentic. The overview calls this framework the successor of AutoGen and Semantic Kernel, names MCP clients, and lists evaluation services among the integrations. It does not, on the page read today, describe a hosted process that runs your agent the way Agent Server or AMP does. You run the process. Cosmos DB is the durable database.

Mastra's storage page is the one that matters for a restart. It keeps workflow snapshots, traces, and eval results after the process exits, and it says the default in-memory store does not. The workflow page covers suspend, resume, and a restart from the last active step. The deploy page says a Mastra server runs on any Node host, names a Mastra platform Server as a production target, and ships optional deployers for Vercel, Netlify, and Cloudflare. MCP is both MCPClient and MCPServer, on revision 2026-07-28.

Pydantic AI's durable-execution page is the widest engine list in this set. Agents run durably on Temporal, DBOS, Prefect, Restate, and further engines the page names, nine in total. The page's distinction is worth keeping: an engine journals model requests and tool calls so a crash mid-step can replay. Saving a run and resuming it later does not require that engine. StepPersistence snapshots settled steps. The same page says an engine wraps MCP server calls, which is a client. This pass did not re-fetch Pydantic's MCP server page. That cell stays on the November 13 matrix.

smolagents is the stale tag that is still a reasonable example. v1.26.0 was published May 29, 2026. The tools page for that version loads MCP tools with MCPClient. The multi-agent page is a notebook of agents sharing a web task. Neither page documents a checkpointer. If the run has to survive a kill, this is the wrong row. 30,360 stars and 203 contributors say the same thing in two directions: people starred it, and the contributor list is the shortest here.

LlamaIndex is the opposite warning. v0.14.26 published November 13, and the repository leads this set in contributors, at 476. The durable-workflows page says there is no built-in checkpointer to enable. You snapshot context when a step finishes and pass that snapshot back to run() after a crash. Resume is at-least-once: a step that was mid-flight runs again, so the side effect has to be safe to repeat. The page's alternative is a runtime plugin. It names DBOS, which journals steps, and WorkflowServer, which can run that plugin. Do not describe LlamaIndex as "it checkpoints" without that sentence.

AG2 is easy to mis-file as AutoGen. The README says that as of v1.0, ConversableAgent, GroupChat, and import autogen live in ag2ai/ag2-classic. This repository is pip install ag2. It had 5,063 stars and release v.1.3.1 from November 10. The README describes a network of channels and persistent knowledge in the harness. Persistent knowledge is not a workflow checkpoint. This note did not fetch a crash-resume page, so that cell stays empty.

AutoGen's row is the one the star sort gets wrong. The README says the project has been archived and is no longer maintained, that Microsoft is not providing bug fixes, security fixes, new features, or support, and that new users should start with Microsoft Agent Framework. The API, the same morning, returned archived: false and 61,809 stars. Both sentences are true. The November 9 note is the place that walks the commit that changed the README. On that morning the count was 61,753 and Agent Framework's latest tag was python-1.25.0. Today the AutoGen count is 61,809 and Agent Framework's latest tag is python-1.26.0, published 2026-11-13T08:28:16Z. The successor gained a GitHub release. The predecessor did not.

What moved since the November 13 matrix?

Three clocks, and nothing else this note is willing to call a change.

The protocol note said PyPI already had agent-framework 1.26.0 and that GitHub had no release tag python-1.26.0. At 03:46 UTC on November 14, GET /repos/microsoft/agent-framework/releases/latest returned python-1.26.0, published 2026-11-13T08:28:16Z. The tag exists. This note did not read the release body, so it does not say what the tag contains.

That note also said npm @mastra/core was 1.83.0 while GitHub's latest release was still 1.82.0. GitHub's latest release is now @mastra/core@1.83.0, published 2026-11-13T14:32:29Z. Registry and tag match on the version number. This note did not re-query npm.

LangGraph was 1.2.21 on that note, uploaded November 12. Release 1.2.22 published at 01:43 UTC on November 14. The body says the changes since 1.2.21 include skipping message conversion for plain messages, filtering pending writes, and streaming a child graph on its own thread. It also tags checkpoint 4.3.2 and checkpoint-postgres 3.2.1. Quote the tag. Do not call it a feature release.

What this means when you pick

Builders. Sort by the crash. If a killed process must continue the same graph, start with LangGraph and a checkpointer that is not MemorySaver, or with Pydantic AI and one named engine. If the team writes TypeScript, Mastra's storage page is the test: the run has to resume after the process exits, on a store that is not the default memory. If the team is on Google Cloud and wants the eval in the same repo as the agent, ADK's deploy page and eval page are a pair. If the design is roles and a product manager will touch the workflow, CrewAI's free 50 executions are a real trial, and Enterprise is a conversation, not a sticker. If you are already on AutoGen, the README and the Learn overview agree on the exit. Agent Framework 1.26.0 is the tag. The 61,809-star repository is the history.

Do not buy MCP support from this table alone when the cell says the November 13 note was the last full read. Client and server are different installs. The matrix is the wire. This page is the job.

Buyers who are not going to import a package should not use this table at all. A personal agent that books, browses, and stops before a payment is a product, not a framework. That reading is ChatGPT agent is retired. Gemini tasks skip the EEA and the UK.

What should you open next?

The counts move. The job does not.

• AI agent framework GitHub stats, November 2026 is the November 5 snapshot of a wider set. This note is 11 repositories, pulled November 14.
• 13 MCP clients, 5 servers, 6 A2A is the protocol matrix. Two of its GitHub tags have since moved. The protocol cells were not re-run.
• LangGraph vs CrewAI vs the OpenAI Agents SDK is the bill for three of these rows, read November 10.
• Is AutoGen deprecated? is the November 9 reading of the README and the archived field. The star count there is 61,753.
• Which AI agent framework do people install? is file requests. A download is not a user, and it is not a star.
• Best AI agent frameworks is the licence and release-tag comparison on the best-of page.
• The stars tracker moves after this page. A number here is the 03:46 UTC pull, not the tracker.