LangGraph vs CrewAI vs AutoGen vs Agents SDK: which to learn for 2027

As a first agent framework for 2027, the OpenAI Agents SDK has a very small set of core primitives, LangGraph suits explicit control over state and control flow, and CrewAI fits teams of role-based agents. AutoGen is no longer a sensible choice for new work: its README says Microsoft archived it and tells new users to start with Microsoft Agent Framework, which reached version 1.0 on 3 April 2026.

How do the four frameworks compare on state, tools and status?

The table compares the four frameworks, plus Microsoft Agent Framework as AutoGen's named successor, on the points a learner meets first. Cells reflect each project's own documentation, README or package page as of 7 November 2026.

LangGraph, CrewAI, the OpenAI Agents SDK and Microsoft Agent Framework are all MIT licensed, so licence does not decide the choice. Each of these Python packages needs Python 3.10 or later.

Is AutoGen still maintained?

No. AutoGen's README states that the project is archived and no longer actively maintained by Microsoft, and that Microsoft provides no bug fixes, security fixes, new features or support.

The README adds that AutoGen was developed for research and experimentation and advises against production use. The README's maintenance banner was updated on 6 April 2026, three days after the Agent Framework 1.0 announcement, and a commit on 16 October 2026 made the archived wording explicit. The latest autogen-agentchat release on PyPI is 0.7.5, from 30 September 2025.

Microsoft Agent Framework is the successor the README names. Its 1.0 announcement says it unifies the enterprise foundations of Semantic Kernel with the orchestrations of AutoGen, and lists sequential, concurrent, handoff, group chat and Magentic-One patterns, all with streaming, checkpointing and human-in-the-loop approvals. A tutorial that imports autogen_agentchat is describing the archived library, and Microsoft publishes a migration guide from AutoGen to Agent Framework.

Can LangGraph be used without LangChain?

Yes. LangGraph's documentation says its examples often use LangChain for model and tool integrations but that LangGraph does not require LangChain, and the README says LangGraph is built by LangChain Inc but can be used without it.

LangGraph is a low-level orchestration framework and runtime for long-running, stateful agents. A workflow is a graph of shared state, nodes and edges, and a node can be ordinary code or a model call. It does not prescribe prompts or agent architecture, so a learner writes more structure by hand than in the higher-level frameworks.

The sample below is a hypothetical refund workflow with one node that pauses for a human decision. LangGraph needs a checkpointer to save state and a thread ID to know which saved run to resume, and Command(resume=True) supplies the human's answer.

from typing import TypedDict
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.graph import END, START, StateGraph
from langgraph.types import Command, interrupt

class State(TypedDict):
    refund: int
    status: str

def approve(state: State):
    ok = interrupt(f"Approve refund of Rs {state['refund']}?")
    return {"status": "paid" if ok else "rejected"}

graph = StateGraph(State)
graph.add_node("approve", approve)
graph.add_edge(START, "approve")
graph.add_edge("approve", END)
app = graph.compile(checkpointer=InMemorySaver())

config = {"configurable": {"thread_id": "t1"}}
paused = app.invoke({"refund": 500, "status": "pending"}, config)
print(paused["__interrupt__"][0].value)
print(app.invoke(Command(resume=True), config))

Output from running it with Python 3.12 and langgraph 1.2.20:

Approve refund of Rs 500?
{'refund': 500, 'status': 'paid'}

The first call stops at interrupt() and returns the question. The saved state waits indefinitely, and the second call resumes the same thread with the answer. Durable execution and human-in-the-loop are two of the capabilities LangGraph's overview names.

How do CrewAI and the OpenAI Agents SDK differ from LangGraph?

CrewAI is an open-source Python framework with two parts: Crews, which are teams of role-based agents that collaborate, and Flows, which are event-driven workflows with state and branching. Its FAQ calls it a standalone framework with its own primitives, so it does not sit on top of LangChain.

Human input is a task-level flag: with human_input set, the agent asks the user before giving its final answer. CrewAI requires Python 3.10 or later but below 3.14, which matters on a machine that already has a newer Python.

The OpenAI Agents SDK is described by its own docs as having a very small set of primitives: agents, agents as tools or handoffs, and guardrails. Tracing is built in and enabled by default. A handoff is shown to the model as a tool, such as transfer_to_refund_agent, and a tool can be marked as needing approval so the run pauses until a person approves.

Despite the name, the SDK is provider-agnostic: the README lists the OpenAI Responses and Chat Completions APIs and 100+ other LLMs. The Python package was at version 0.24.0 on 15 October 2026, a number below 1.0, so the API can still change.

Which framework should a beginner learn first?

The answer depends on the criterion. For the fewest concepts before a first working agent, start with the OpenAI Agents SDK, whose docs describe a very small set of primitives. Other situations point elsewhere:

• For explicit control over state, branching and pausing for approval, LangGraph fits better, because the workflow is a graph of state, nodes and edges.
• For a quick prototype that mirrors a team of specialists, CrewAI's Crews model matches that picture and needs only Python.
• For a .NET or Azure-centred stack, Microsoft Agent Framework supports Python and .NET with stable 1.0 APIs.
• For reading older multi-agent code, AutoGen material still explains ideas such as group chats, but no new project should depend on it.

Before any of them, build the loop once without a framework. OpenAI's docs advise calling the Responses API directly when you want to own the loop, tool dispatch and state, and Anthropic's guidance of 19 December 2024 suggests starting with LLM APIs directly because many patterns need only a few lines of code.

Tools can also come from servers that speak the Model Context Protocol, which the OpenAI Agents SDK, Microsoft Agent Framework, AutoGen and CrewAI all document. The free agentic AI courses include a LangGraph course and a course that covers smolagents, LlamaIndex and LangGraph side by side.

For usage context, Stack Overflow's 2026 Developer Survey, with 30,903 responses, reports LangChain at 30% and the OpenAI Agents SDK at 24% for orchestration. Usage is not quality, and the survey does not say which framework suits a first project.

What are the limits of choosing a framework?

Frameworks add layers. Anthropic warned in December 2024 that they often obscure the underlying prompts and responses and make debugging harder, and the same page now notes that much of the tooling it described has changed. A guide written a year ago can describe an API that has since changed, and AutoGen tutorials now describe an archived library.

Young version numbers carry risk. The OpenAI Agents SDK is at 0.24.0, so pin the version in a requirements file. Setup has limits too: CrewAI stops below Python 3.14, and every one of these packages needs Python 3.10 or later.

No framework tests an agent on real cases or limits what its tools can do. Those jobs stay with the developer, whichever library runs the loop. For readers still sorting out the terms, AI agents vs agentic AI vs generative AI explains how they differ, and top agentic AI tools and frameworks lists wider tooling.

Where does framework knowledge fit in an AI and ML programme?

Certification in Artificial Intelligence and Machine Learning is an online programme from Vishlesan i-Hub, IIT Patna, delivered on Masai's online platform, that covers Python, machine learning, deep learning, generative AI and LLMs, agentic systems and deploying models to production. Its modules include Generative AI, LLM Engineering and Agentic Systems, and Production Deployment, MLOps and Capstone, the stage at which a framework choice becomes a practical decision.

Frequently asked questions

Which AI agent framework is best for beginners?

No framework is best on every measure. By the criterion of fewest core concepts, the OpenAI Agents SDK, whose docs list a very small set of primitives, is the easiest start. By the criterion of explicit state and pausing, LangGraph suits better. Building one agent loop by hand before choosing either makes the abstractions easier to judge.

Is AutoGen still maintained?

No. AutoGen's README says Microsoft has archived it and provides no bug fixes, security fixes, new features or support. The last package release on PyPI, 0.7.5, dates from 30 September 2025. Microsoft Agent Framework, which reached 1.0 on 3 April 2026, is the successor the README names for new users.

Can LangGraph be used without LangChain?

Yes. LangGraph's documentation says it does not require LangChain, and the README says it is built by LangChain Inc but can be used without it. Many tutorials still import LangChain for model and tool integrations, so learners meet both libraries even when the LangGraph code itself stands alone.

Does the OpenAI Agents SDK only work with OpenAI models?

No. The README describes the SDK as provider-agnostic, supporting the OpenAI Responses and Chat Completions APIs and 100+ other LLMs. The Python package needs Python 3.10 or later, and a separate repository covers JavaScript and TypeScript. The Python package is still at a 0.x version, 0.24.0, so pinning it is wise.

Do these frameworks support human approval before an action?

Yes, each in its own way. LangGraph pauses with interrupt() and resumes with Command. The OpenAI Agents SDK lets a tool require approval through needs_approval. CrewAI has a human_input flag on tasks, and Microsoft Agent Framework lists human-in-the-loop approvals for its orchestration patterns.

Next step

Pick one framework, pin its version and build the same small refund assistant in it before trying a second. The Certification in Artificial Intelligence and Machine Learning page lists the modules on agentic systems and production deployment for readers who want a structured route, and the free resources linked above cover the same ground for self-study.