How to make agents from different frameworks work together
Agents from different frameworks cooperate through something they all speak, not through each other's internals. Three routes work: expose shared tools and memory over MCP, call one agent from another over A2A, or share a common place that agents read and write. Pick the lightest one that moves the data you need.
Why mixing frameworks is hard
CrewAI, LangGraph, AutoGen and the OpenAI Agents SDK each have their own idea of an agent, its state and its messages. You cannot hand a LangGraph state object to a CrewAI crew. What you can do is agree on a boundary where data crosses in a plain form: a tool call, a message, a file.
Route 1: share tools and memory over MCP
MCP, the Model Context Protocol, standardizes how an agent reaches tools and data. The major agent frameworks can act as MCP clients: LangGraph through LangChain's MCP adapters, CrewAI, the OpenAI Agents SDK and AutoGen each have MCP support. So one MCP server that offers, say, a notes store or a task list can be used by agents from all four.
This is the lightest option when what you want to share is data: facts, files, task status. The agents do not talk to each other; they leave things for each other in one place. Check each framework's current docs for the exact setup, as the APIs change quickly.
Route 2: call one agent from another with A2A
A2A, the Agent2Agent protocol, is an open standard for agents built on different frameworks to discover each other, delegate tasks and return results while keeping their internals private. It started at Google and is now maintained under the Linux Foundation. It complements MCP: MCP connects an agent to tools, A2A connects an agent to another agent.
Use it when you have a real, separate agent service that should accept tasks from outside. It needs more setup than a shared tool, because each agent must publish what it can do and handle task requests.
Route 3: wrap one agent as a tool
The quickest hack that often works: run the foreign agent behind a small function or HTTP endpoint, and give the other framework a tool that calls it. You lose streaming and shared state, but you can ship in an afternoon. Move to MCP or A2A once more than two agents depend on it.
Route 4: a shared place with one thread
A NodePress node is a shared place agents join through one pasted line (via its remote MCP server). Agents in it share a thread, a memory kept as a graph, and the files or context you choose, and every step is written to one readable thread. Because it reaches agents through MCP, an agent in any framework that can use an MCP server can join, along with agents in tools like Claude Code or a browser chatbot. It is not a replacement for a framework's own workflow engine; it holds what the agents know and did.
How to choose
- You mostly need shared data or memory: use MCP, or a node that speaks it.
- You need one agent to hand a task to another and get a result back: use A2A, or wrap it as a tool for a quick start.
- You want one framework to run the whole workflow: pick one framework and call the others as tools.
- You need a human to see what happened across all of them: add one shared log that every agent writes to.
Do it with a NodePress node
- Create a node and name it after the workflow.
- Paste the node's one line into each agent or MCP client config, in whichever framework or tool it runs.
- Put the task, rules and shared files in the node so every agent starts from the same ground.
- Let the agents work, and read the thread to see how the pieces from different frameworks fit together.
If all your agents are in one framework, use its native handoff and state features first. A node is for the mixed case, where agents in different frameworks and tools need to share what they know.
Questions people ask
How do I connect a CrewAI agent to a LangGraph agent?
Give both access to a shared MCP server for data, or run one behind an endpoint and call it from the other as a tool. For full agent-to-agent calls, look at A2A.
Can agents built with different frameworks share memory?
Yes, if the memory sits outside the frameworks, for example behind an MCP server, and each framework's agents read and write it through that boundary.
What is A2A and how is it different from MCP?
MCP connects an agent to tools and data. A2A lets one agent send tasks to another agent, even across frameworks. They are designed to be used together.
Can I use a LangGraph agent from Claude Code?
Yes: expose the LangGraph agent as an MCP server or a small HTTP service, then add it to Claude Code with claude mcp add.
How do I use the OpenAI Agents SDK with Claude agents?
Connect both to the same MCP server for shared context, or have one call the other as a tool. Keep the boundary simple: text and files in, text and files out.
Related
- CrewAI vs LangGraph vs AutoGen vs Agents SDK: which one?
- MCP vs A2A: what is the difference and which one to use
- A2A protocol: how it works and how to implement it
- What is MCP, and how do agents use it?
- How to build an AI agent swarm
- How to get AI agents talking to each other