What Is the Most Popular AI Agent Framework?

The most popular AI agent framework in 2026 is LangChain, together with its orchestration layer LangGraph. It leads on all three public measures we checked: it is the most used agent framework in the 2026 Stack Overflow Developer Survey (30.2% of 4,400 respondents), the most downloaded agent library on PyPI (about 170 million installs in the last month), and the most starred agent framework repo on GitHub (about 148,000 stars). Second place depends on what you count: the OpenAI Agents SDK is second by survey use, while AutoGen and CrewAI are second and third by stars.

Popular is not the same as right for your project, so this guide shows the numbers, explains what each one actually measures, and gives a short script so you can rerun the check yourself. If you want a pick by use case instead of a popularity ranking, read our best AI agent framework for Python guide or the best TypeScript agent frameworks roundup.

The most popular AI agent framework, by three measures

We pulled every number below on October 23, 2026 from public, first party sources: the Stack Overflow 2026 survey data, the PyPI Stats API (last 30 days, known mirrors excluded), the npm downloads API (September 19 to October 18) and the GitHub REST API for stars.

A few numbers outside the table help the picture. The Claude Agent SDK, which the survey did not list, had 29.8M PyPI downloads (claude-agent-sdk) and 47.9M npm downloads (@anthropic-ai/claude-agent-sdk). LlamaIndex has 52.6k stars, DSPy 38.8k and Agno 42.9k. Microsoft Agent Framework, the successor to AutoGen, has 14.3k stars and about 1.0M monthly downloads of agent-framework-core.

Measure 1: what developers say they use

The cleanest signal is the 2026 Stack Overflow Developer Survey, which asked: "Have you used any of the following tools for AI agent orchestration or agent frameworks in the past year?" Of the 4,400 people who answered, 30.2% picked LangChain, 23.8% the OpenAI Agents SDK, 18.9% "Terminal Use", 18.8% LangGraph, 17.4% Llama Stack and 14.0% Google ADK. CrewAI sat at 5.5% and AutoGen at 4.5%.

Two things stand out. First, LangChain's share slipped a little: in the 2025 survey, 33% of developers building agents used it. Second, the OpenAI Agents SDK, first published on PyPI in March 2025, is already in second place. Stack Overflow's own note on the 2026 result is that the "Terminal Use" answer suggests many developers favor direct, lightweight orchestration over dedicated frameworks. That matches what we see with coding agents: plenty of people now build agents as a plain loop around a model API, which is the approach in our build an AI agent from scratch tutorial.

Survey shares are self reported and respondents could pick several tools, so the percentages add up to well over 100%.

Measure 2: package downloads

Downloads show how often a library is installed. LangChain wins on PyPI by a wide margin, and langchain-core alone pulled 144.8M installs in the same 30 days. LangGraph is a clear second among dedicated agent frameworks at 44.8M.

Downloads also produce some surprises. Strands Agents, AWS's open source agent SDK, logged 37.1M PyPI downloads while only 4.0% of survey respondents said they use it. On npm, the Vercel AI SDK's ai package (118.6M) dwarfs every agent framework because it is the general toolkit many JavaScript apps use to call models at all, agent or not.

Treat downloads with care for three reasons:

• Automation counts. Every CI run and every fresh container that installs a package counts as a download. A project with a big test matrix can generate thousands a day.
• Dependencies count. When another popular package depends on a framework, its installs pull the framework in too, whether or not anyone writes code against it.
• Package splits change totals. Pydantic AI ships a full pydantic-ai package (5.4M) and a lighter pydantic-ai-slim (23.9M). Microsoft Agent Framework splits into agent-framework and agent-framework-core. Compare like with like.

Measure 3: GitHub stars

Stars measure attention, not use. LangChain leads with 148.3k. AutoGen (61.5k) and CrewAI (60.0k) follow, then LlamaIndex (52.6k), LangGraph (43.8k) and Agno (42.9k).

AutoGen is the clearest example of why stars mislead. Its GitHub README now marks the project as archived: Microsoft provides no bug fixes, security fixes or new features, advises against production use, and points new users to Microsoft Agent Framework with a migration guide. Its stars stay high because they never expire. Its 0.4M monthly downloads tell you more about where it is now.

Stars also favor older projects. LangChain's repo dates from October 2022, while the OpenAI Agents SDK and Google ADK repos were created in 2025.

So which framework is the most popular?

Putting the three together:

1. LangChain plus LangGraph is the most popular AI agent framework family by every measure. LangChain 1.x lists LangGraph as a required dependency, so every LangChain install brings LangGraph with it. Counted as one ecosystem, nothing else is close.
2. OpenAI Agents SDK is the fastest riser: second in the survey only a year and a half after its first PyPI release in March 2025, with 12.4M monthly downloads.
3. Google ADK has real use too: 14.0% in the survey and 10.1M monthly PyPI downloads.
4. CrewAI has a large fan base on GitHub but a much smaller share of measured use.
5. Vercel AI SDK and Mastra lead the TypeScript side, where the Python numbers do not apply.

How to choose an AI agent framework

Popularity is a fair tiebreaker. A big community means more examples, more answered questions and more integrations. It should not be your first filter. Ask these questions in this order:

• Language. Python teams usually pick between LangGraph, the OpenAI Agents SDK, Pydantic AI, CrewAI and Google ADK. TypeScript teams usually pick between the Vercel AI SDK, Mastra and LangGraph.js.
• How much control you need. If your agent needs branches, retries, durable state and a human approval step, choose a graph runtime such as LangGraph. Our LangChain vs LangGraph guide explains where one stops and the other starts.
• Model lock in. The OpenAI Agents SDK and Google ADK are built around their vendors' models, though both document LiteLLM support for other providers. LangChain and Pydantic AI are provider neutral by design.
• Maintenance. Check the last release date and the README before you adopt anything. An archived project like AutoGen can still top a star chart.
• Whether you need a framework at all. A config driven runtime such as Docker Agent or a small hand written loop is often enough. Our agent harness explainer covers what a framework actually adds.

For a side by side of the three most common Python picks, see LangGraph vs CrewAI vs OpenAI Agents SDK.

Check the numbers yourself

These rankings move monthly. The script below uses only the Python standard library and three public endpoints that need no API key. We ran it on Python 3 on October 23, 2026, and it printed the PyPI, npm and star figures shown in the table above.

import json, urllib.request

PYPI = ["langchain", "langgraph", "openai-agents", "google-adk", "crewai"]
NPM = ["@langchain/langgraph", "@openai/agents", "@mastra/core"]
REPOS = ["langchain-ai/langchain", "crewAIInc/crewAI", "openai/openai-agents-python"]

def get(url):
    req = urllib.request.Request(url, headers={"User-Agent": "popularity-check"})
    with urllib.request.urlopen(req, timeout=20) as r:
        return json.load(r)

for p in PYPI:
    d = get(f"https://pypistats.org/api/packages/{p}/recent?period=month")
    print(f"pypi  {p:28} {d['data']['last_month']:>12,}")
for p in NPM:
    d = get(f"https://api.npmjs.org/downloads/point/last-month/{p}")
    print(f"npm   {p:28} {d['downloads']:>12,}")
for r in REPOS:
    d = get(f"https://api.github.com/repos/{r}")
    print(f"stars {r:28} {d['stargazers_count']:>12,}")

PyPI Stats updates once a day and asks people to cache results rather than poll. Unauthenticated GitHub API calls are rate limited, so keep the repo list short or pass a token.

FAQ

Is LangChain still the most popular AI agent framework?

Yes. In October 2026 it leads the Stack Overflow survey, PyPI downloads and GitHub stars. Its survey share did fall from 33% in 2025 to 30.2% in 2026 while newer SDKs grew.

What is a multi agent AI framework?

A multiagent framework lets several agents, each with its own instructions and tools, hand work to each other. CrewAI organizes agents into crews with roles, the OpenAI Agents SDK uses handoffs, and LangGraph lets you wire agents as nodes in a graph. Many tasks work fine with a single agent, so add more only when the jobs really differ.

Is AutoGen still maintained?

No. Microsoft has archived AutoGen and recommends Microsoft Agent Framework for new projects, with a published migration guide. AutoGen gets no security fixes.

What is Mosaic AI Agent Framework?

It is Databricks' set of tools for building and deploying agents on its platform. The current Databricks docs present this area as Agent Bricks, with an agent runtime, sandbox, memory, MCP server support and MLflow tracing. It hosts agents written with other frameworks rather than competing with LangChain on the same layer.

Which AI agent framework is most popular for TypeScript?

By npm downloads, the Vercel AI SDK leads by a wide margin, followed by LangGraph.js, the OpenAI Agents SDK and Mastra. Our TypeScript framework guide compares them.

Do the popular frameworks support MCP?

The leaders do. The OpenAI Agents SDK, Google ADK and CrewAI each have an MCP section in their official docs, and LangChain loads MCP tools through its langchain-mcp-adapters package. Check the docs of any smaller framework before you assume support. See MCP vs A2A for how the protocol fits next to agent to agent messaging.