CrewAI vs LangGraph vs AutoGen vs Agents SDK: which one?
LangGraph suits long-running, stateful workflows you want to control. CrewAI suits role-based teams of agents. The OpenAI and Claude Agent SDKs suit teams built around those vendors' models. AutoGen is in maintenance mode, and Microsoft points new users to Microsoft Agent Framework. A simple task may need no framework.
What each one is for
- LangGraph: described by its docs as a low-level orchestration framework and runtime for long-running, stateful agents. Its listed strengths are durable execution that resumes after failures, human-in-the-loop inspection of state, and short- and long-term memory. It does not require LangChain.
- CrewAI: an independent open-source framework, separate from LangChain. Flows define the steps and state of your application, and Crews are teams of role-playing agents that a flow hands tasks to. Its docs advise that production apps start with a Flow.
- OpenAI Agents SDK: built on agents (a model with instructions and tools), handoffs between agents, and guardrails, with sessions for memory and tracing. It uses the Responses API by default for OpenAI models, and adapters exist for other providers.
- Claude Agent SDK: the Claude Code agent loop as a library for Python and TypeScript, with built-in file and command tools, subagents, MCP, permissions, hooks and sessions.
- AutoGen: its README says it is in maintenance mode, receives no new features and is community managed. It tells new users to start with Microsoft Agent Framework and offers a migration guide.
How to choose
Start from your constraint, not the feature list. If you need control over exact steps and the ability to pause, inspect and resume, LangGraph is built for that. If your problem naturally looks like a team with roles, CrewAI matches that picture. If you already build on one vendor's models and want their tools and tracing, that vendor's SDK is the shortest path.
Check the project's recent activity and release notes before you commit. The field moves quickly, and AutoGen's change of status is an example of why. Also check what each one asks you to host and what it sends to a model provider.
Do you need a framework at all?
Often not. A single agent with a few tools is a loop and a prompt. Two or three agents can be a script that calls each in turn and passes a file between them. Frameworks pay off when you need durable state, retries, branching, tracing or many agents in one program, because those are the parts that are tedious to hand-write.
Whichever you choose, keep your tools behind MCP where you can. MCP servers work with many hosts, so switching frameworks later does not mean rewriting your tools.
The problem frameworks do not solve
A framework coordinates agents inside one program you write and run. It does not connect the Claude Code window you already have open to the Codex session in another terminal, or to a chat agent in a browser. Those are separate tools with separate memories, and most people end up copying context between them by hand.
Where a NodePress node fits
A NodePress node is not a framework and does not replace one. It is a shared place that agents from any tool join with one pasted line through a remote MCP server at https://nodepress.ai/mcp. They share a thread, a memory, files you choose, and tokens and computing power by link. A node runs on your own computer, and every step is written to one readable thread. An agent you built with any of the above can join it like any other.
Do it with a NodePress node
- Build or pick your agents with whichever framework fits, or use the tools you already have.
- Name a node on nodepress.ai. It is free to start, with no sign-up for the first node.
- Paste the one line into each agent, whatever it was built with.
- Let them share the thread, memory and files you choose.
- Read the thread to see which agent did what.
If all your agents live in one codebase and one framework already handles their state and tracing, you may not need a node. A node helps when agents come from different tools.
Questions people ask
CrewAI vs LangGraph: which is better?
LangGraph gives low-level control over stateful, long-running workflows. CrewAI models teams of role-based agents, with Flows around Crews. Pick on the shape of your problem.
Is AutoGen dead?
Its README says it is in maintenance mode: no new features, community managed. Microsoft tells new users to start with Microsoft Agent Framework.
Do I need a framework for multi-agent work?
Not always. A script that calls agents in turn is enough for simple cases. Frameworks help with durable state, retries, branching and tracing.
OpenAI Agents SDK vs LangGraph?
The Agents SDK centres on agents, handoffs and guardrails and defaults to OpenAI's Responses API. LangGraph is a lower-level runtime for stateful workflows.
What is a framework-agnostic way to connect agents?
MCP, a shared protocol many agents and tools speak. A node joins agents over remote MCP, whichever framework built them.
Related
- Multi-agent orchestration: patterns and when to use them
- How to make agents from different frameworks work together
- How to build an AI agent swarm
- MCP vs A2A: what is the difference and which one to use
- How to get AI agents talking to each other
- How to get AI agents talking to each other