AI Agent Frameworks Compared

Every comparison of agent frameworks you will find quotes download counts and speed claims with no date and no source. Those numbers were often stale when published and are certainly stale now.

So this page does something narrower and more useful. The figures below were measured directly from the GitHub API on 15 August 2026 and are stated with that date attached. Check them yourself; they will have moved.

And a warning about what they mean: stars measure attention, not suitability. A framework with four times the stars is not four times better for your problem. The genuinely informative column is the last one.

Framework	Repository	Stars	Last pushed
LangChain	langchain-ai/langchain	144,284	2026-08-15
Dify	langgenius/dify	152,547	2026-08-15
AutoGen	microsoft/autogen	60,436	2026-04-15
CrewAI	crewAIInc/crewAI	57,124	2026-08-15
LlamaIndex	run-llama/llama_index	51,662	2026-08-14
LangGraph	langchain-ai/langgraph	39,750	2026-08-14
smolagents	huggingface/smolagents	28,813	2026-07-21
OpenAI Agents SDK	openai/openai-agents-python	28,661	2026-08-15
Semantic Kernel	microsoft/semantic-kernel	28,452	2026-08-11
Pydantic AI	pydantic/pydantic-ai	19,317	2026-08-15

Start with the thing nobody selling a framework will tell you

You may not need one.

An agent is a loop: call the model, execute the tool it asked for, feed the result back, repeat until done or until a limit stops you. In most languages that is under a hundred lines, and you will understand every one of them.

Anthropic's own guidance in Building Effective Agents says as much:

When building applications with LLMs, we recommend finding the simplest solution possible, and only increasing complexity when needed.

The cost of a framework is not the dependency. It is that when the agent behaves strangely — and it will — you are debugging your logic and someone else's abstraction, through a stack trace that passes through six layers of callback. Frameworks earn their place when they give you something genuinely hard to build: durable state across restarts, checkpointing and resumption, human-in-the-loop interrupts, distributed execution, observability. If you are not using those, you are paying the abstraction cost for a wrapper around a while-loop.

Decide this before you shop. Decide Whether Your Task Actually Needs an Agent and AI agent vs workflow vs single call both come before this page.

The one unambiguous finding

AutoGen is in maintenance mode. Its own README states it plainly:

AutoGen is now in maintenance mode. It will not receive new features or enhancements and is community managed going forward.

New users should start with Microsoft Agent Framework.

The 60,436 stars are real and so is the four-month gap since the last push. This is exactly the failure mode of star-ranked listicles: a project can sit near the top of every comparison for a year after new development moved elsewhere. If you are starting on Microsoft's stack today, the destination is Microsoft Agent Framework, not AutoGen.

What each one is actually for

LangChain is the largest ecosystem, and its real value is breadth of integrations — hundreds of model providers, vector stores, loaders and tools already wrapped. It is also the most criticised, usually for abstraction depth. Reasonable use: you want the integrations. Less reasonable: you want a clean agent loop.

LangGraph is the LangChain team's answer to that criticism, and a genuinely different thing. You define agents as an explicit graph of nodes and edges, with state that persists and can be checkpointed and resumed. That makes it the strongest option when your requirement is durability — long-running processes that must survive a restart, or workflows needing human approval mid-flight. The cost is that you model your problem as a graph, which is real work when your problem is not shaped like one.

CrewAI models a team: agents with roles, delegating to each other. It is the fastest way to a working multi-agent prototype and the abstraction is intuitive. The risk is that role-play is a compelling metaphor which does not always correspond to better output — several crews would perform identically as one well-prompted call, at a fraction of the token cost. See multi-agent systems explained.

OpenAI Agents SDK is deliberately minimal — a small number of primitives (agents, handoffs, guardrails, tools) and very little else. Its strength is that you can read the whole thing. Its constraint is tight coupling to OpenAI's hosted tooling, which is a genuine cost if you expect to switch providers.

Pydantic AI brings the Pydantic model to agents: typed inputs and outputs, validation as a first-class concern. If your pain is unreliable output shape, this addresses it at the framework level rather than the prompt level — related, getting reliable structured output from LLMs.

LlamaIndex started as a data framework and remains strongest there. If the hard part of your problem is ingestion, chunking and retrieval rather than orchestration, it is aimed at you. See what are embeddings and what is a vector database.

Semantic Kernel is Microsoft's enterprise-oriented SDK with first-class C# and .NET support — often the deciding factor, because most of this ecosystem is Python-first.

smolagents is Hugging Face's minimal library, notable for a code-writing agent approach where the model emits Python rather than JSON tool calls.

Dify is a different category despite topping the star count: a self-hostable platform with a visual builder, not a library you import. Comparing it to LangGraph is comparing a product to a package. Worth knowing if what you actually want is a UI for non-engineers.

How to choose

Work down these in order and stop when one decides it:

1. Do you need durable state, checkpointing, or human-in-the-loop interrupts? If yes, that narrows the field hard, and LangGraph is the most developed answer.
2. What language is your production stack? Most of this is Python-first. If you are .NET, that decides it.
3. Are you committed to one model provider? A vendor SDK is simpler if yes and a liability if no.
4. Is your hard problem orchestration or retrieval? They point at different tools.
5. Do you need a GUI for non-engineers? That is a platform, not a library.
6. If none of the above applies — write the loop.

Whatever you pick, the things that actually determine whether your agent works are not framework features. They are tool design, evals, and guardrails: how to design tools your AI agent can actually use, how to write evals for your AI agent, and guardrails and safety for AI agents. A well-instrumented hand-rolled loop beats an unmeasured framework every time.

Sources

• GitHub REST API, repository metadata for each project listed, retrieved 15 August 2026 — star counts and last-push dates
• Microsoft, AutoGen — README notice placing the project in maintenance mode and directing new users to Microsoft Agent Framework
• Anthropic, Building Effective Agents, 19 December 2024 — find the simplest solution possible, and increase complexity only when needed