3,165 GitHub releases in 30 days, or 407 days of silence

3,165. That is how many non-draft GitHub releases vercel/ai published in the 30 days before 03:42 UTC on November 11, 2026.

The other 23 repositories in this count published 198 in that same window. 3,165 divided by 198 is 15.98. In the 60-day window the Vercel repository published 5,991. microsoft/autogen published zero in both. Its latest tag, python-v0.7.5, is dated September 30, 2025, which is 407 calendar days before the pull date.

ag2ai/ag2 published four releases in the 30 days and nine in the 60. The repository description calls AG2 formerly AutoGen. The release list agrees with that split: the repository under the old name is quiet, and the repository under the new name is not. What the README says about maintenance is Is AutoGen deprecated?. This note does not re-argue it. It counts tags.

The November 5 snapshot is one latest tag per repository, for 19 of these rows. Best AI agent frameworks compares licences and those tags. Neither page is a cadence. This one is.

The 30-day count and the 60-day count

A release is in the 30-day window when published_at is on or after 2026-10-12T03:42:41Z, and in the 60-day window when it is on or after 2026-09-12T03:42:41Z. Days are calendar days from that timestamp’s UTC date to 2026-11-11. Prereleases are inside the counts. The Pre column is how many of the 30-day releases were marked prerelease. Vercel’s later pages were not split that way. The first 1,500, covering October 27 through November 10, had zero prereleases.

Add the 30-day column without Vercel and you get 198. That is the audit of the ratio.

Four repositories are empty in both windows: AutoGen, Letta, smolagents, and CAMEL. LlamaIndex is empty in 30 days and has one release in 60, v0.14.25 on September 21, 51 days before the pull. CAMEL’s 69 days belong to the prerelease v0.2.91a7. The newest tag that is not a prerelease is v0.2.90, published March 22, 2026, 234 days before the pull, and it sits outside the table’s latest-tag cell because the cell is the newest release GitHub returned.

A package tag is not a library version

Vercel’s latest tag is @ai-sdk/zai@3.0.52, published 2026-11-10T20:21:32Z. The first 1,500 releases, the ones stored with their tags, used 86 tag prefixes. @ai-sdk/vue, @ai-sdk/react, @ai-sdk/llamaindex, @ai-sdk/langchain, and ai each appeared 32 times in that slice. A changeset monorepo that publishes one tag per package will post a hundred releases on a day when a single-package library posts one. 3,165 is a real count of release objects. It is not 3,165 copies of one SDK.

LangChain’s 32 and LangGraph’s 30 are the same shape, at a smaller scale. LangGraph’s 30-day tags include cli, sdk, checkpoint, checkpointsqlite, checkpointpostgres, and prebuilt, plus six tags whose names are only the version: 1.2.15 on October 12, 1.2.16 on October 16, 1.2.17 on October 24, 1.2.18 on October 26, 1.2.19 on October 30, and 1.2.20 on November 5 at 02:20:53Z. That last timestamp is the one in the November 5 snapshot. The core series has not moved. The newest release on the repository, cli==0.4.42, was published November 6. A page that prints only the latest tag now names the CLI.

Strands was left out of that November 5 snapshot because strands-agents/sdk-python redirected. It still does. The 301 target’s release list is strands-agents/harness-sdk: 25 releases in 30 days, 49 in 60, latest harness-cli/v0.3.0 on November 10. Across the 60 days the 49 tags use six prefixes: python 13, harness-cli 11, harness-python 9, harness-typescript 8, typescript 7, and mcp 1.

Google’s three ADK repositories do not share a latest tag. google/adk-python published 3 releases in 30 days, v2.12.0 on October 13, v2.13.0 on October 23, and v2.14.0 on November 7. google/adk-js published 8, which are two versions, 2.4.0 and 2.3.0, each cut into four package tags in the same minute. The newest is integrations-v2.4.0 on October 28, 14 days before the pull, and adk-v2.4.0 is in that same minute. google/adk-go published 5. The newest is plugin/agentanalytics/v0.1.0 on November 5. The newest v2 tag is v2.7.1 on November 4, seven days before the pull. The window also contains v1.9.0 on October 15. Two major lines, one releases list.

The rows that are one line of tags are the ones you can compare to each other. Mastra published 7, and all 7 are @mastra/core, from 1.76.0 on October 13 to 1.82.0 on November 4. CrewAI published 12, from 1.15.28 on October 14 to 1.15.39 on November 9, none of them prereleases. The Claude Python agent SDK published 15, from the v0.2 series, latest v0.2.180 on November 9. Pydantic AI published 15, but 12 are v2 and 3 are v1, latest v2.67.0 on November 10 at 23:11 UTC, one calendar day before the pull. Microsoft Agent Framework’s 8 are 4 Python tags (python-1.22.0 through python-1.25.0) and 4 .NET tags (dotnet-1.25.0 through dotnet-1.28.0).

OpenAI’s two SDKs are easy to misread off the latest tag, because both latest tags are alphas. openai/openai-agents-python published 3 releases in 30 days: stable v0.24.0 on October 15, then v0.24.0-alpha.6 on November 3 and v0.24.0-alpha.7 on November 8. The stable tag is 27 days old. The alpha is 3. openai/openai-agents-js published 7: five package tags at 0.21.0 in the same minute on October 20, including @openai/agents@0.21.0, plus v0.22.0-alpha.1 and v0.22.0-alpha.2. Five of the seven are one publish.

Haystack’s 9 in 30 days include 5 prereleases. The latest, v3.5.0 on November 10, is not one of them. Agno published 10, latest v3.1.12 on October 29, 13 days before the pull. DSPy published 2, latest 3.5.0 on November 10. Semantic Kernel published 3, latest dotnet-1.82.0 on November 7.

Four repositories did not ship, and one shipped once

smolagents v1.26.0 was published 2026-05-29T05:09:47Z. That is 166 days. Letta 0.16.8 was published 2026-05-14T17:14:24Z. That is 181 days. LlamaIndex’s only release inside 60 days is v0.14.25 on September 21. AutoGen’s gap is the long one, 407 days, and it is the gap a latest-release cell hides if the cell is allowed to look like “current” merely because the tag exists.

AG2’s four 30-day tags are v1.2.1 on October 21, v1.2.2 on October 27, v1.3.0 on November 4, and v.1.3.1 on November 10. The tag spelling is GitHub’s, including the extra dot. v1.2.0 on October 11 falls just outside the 30-day cutoff and inside the 60.

Quiet is not the same fact as unused. The install census and the downloads tracker count file requests. The stars tracker counts bookmarks. A library can be quiet on tags and loud on downloads, or the reverse. This note does not re-sum those series.

What the notes mark as breaking

Most of the 198 releases do not label a break. These do, in the notes read for this pull.

Google ADK Python v2.14.0, published November 7, has a Breaking changes section. FunctionTool parameters annotated Enum, Optional[Enum], or list[Enum] now receive enum members instead of the raw strings the model sent. The note says string comparison, string-keyed lookups, and json.dumps on a plain Enum no longer work. v2.13.0, published October 23, also has that heading. Telemetry defaults to the OpenTelemetry GenAI v2 schema. The legacy schema stays available by setting ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36.

Microsoft Agent Framework python-1.25.0, published November 7, tells a Python 3.10 environment to move to 3.11 or later before installing. The same notes mark as breaking a bound on MCP tool and prompt discovery, 100 pages and 1,000 items per catalog, and a default 30-second timeout on framework-owned requests. dotnet-1.28.0, published November 6, marks as breaking the removal of A2AServerRegistrationOptions.AgentRunMode. The built-in handler returns a message rather than a task.

Pydantic AI v2.62.0, published October 27, says GraphBuilder broadcast branches now start in .to() order instead of reversed, which can break replay of a durable workflow that was paused across the upgrade. The newer v2.67.0 notes use “Break” as a verb about chat-template special tokens. That line is a behavior description, and this page does not promote it to a breaking-change heading the notes did not use.

AG2 v1.1.1, published September 29, is inside the 60-day window and outside the 30. The notes head a section “Breaking: Bedrock runs on aiobotocore.” BedrockConfig uses a native async client instead of wrapping synchronous boto3.

CrewAI 1.15.34, published November 1, has a Breaking Changes heading that documents the removal of the 1.15.29 agent executor and points at migration notes. 1.15.29 itself was published October 23, inside the 30-day window. The sentence above is what 1.15.34’s notes say. It is not a second reading of 1.15.29’s body.

Haystack v3.5.0 does not use that heading for the splitter change, and the change is still a pin risk. RecursiveDocumentSplitter in word mode no longer counts a whitespace-only piece as a word. The notes say the chunk boundaries move for text that contains page breaks, and they tell you to re-split and re-index if you rely on stable boundaries.

A search that only looks for the letters b-r-e-a-k will also catch “breakpoint” in a LangGraph fix and “breaking restore” in a CrewAI checkpoint note. Those are not in the list above.

What this means if you pin a framework this week

Read the tag shape before you read the count. If you import langgraph, the core tag to pin against is still 1.2.20 from the morning of November 5, not cli==0.4.42. If you import the Vercel AI SDK, 3,165 is the noise of a package publish, and the tag you pin is the package you import, not @ai-sdk/zai@3.0.52 unless that is the package. If you import @mastra/core, seven releases in 30 days, 1.76.0 through 1.82.0, is the cadence of that package.

If you are choosing between the AutoGen repository and AG2, the releases API already separated them. Zero releases in 407 days on microsoft/autogen. Four in 30 days on ag2ai/ag2, latest v.1.3.1 on November 10. The Bedrock client break in the 60-day window is v1.1.1 on September 29, not the November tag.

If you are about to upgrade Google ADK Python to v2.14.0 or Microsoft Agent Framework to python-1.25.0, both published November 7, read the enum-member change and the Python 3.10 drop before you treat the bump as a patch. Those are the two notes in this set that will fail a call or an install, rather than only adding a tag.

Do not treat smolagents’ 166 days, or LlamaIndex’s single September release, as a quality score. Treat them as a maintenance fact, and go to the download table if the question is who installs the library. The November 5 page remains the star, fork, and licence table. This page is the count that table’s latest-release cell cannot make.

The install pins and behavior notes for browser-use 1.0.2, Pydantic AI 2.67.0, AG2 1.3.1, and FastMCP 4.1.8, read the morning of November 11, are AI agent framework release November 2026. browser-use is not one of the 24 repositories in this count.