One of the issues that I am facing with n8n agents is multi-agent systems.
There are all these video tutorials that demonstrate how to make “multi-agent systems” online but upon further inspection you see that the systems are very simplistic. When the parent agent calls the child agent, the child agent performs a single action and reports the results back to the parent agent.
Contrast this with LangGraph’s swarm framework where agents can truly hand off tasks to each other while maintaining memory state. Do you think that we will be able to build agent systems with this framework on n8n soon. My clients and myself want to build single interface multiagent systems for our businesses, I wanted some thoughts on whether you think I should go learn Lang Graph, or just wait for more n8n updates?
These are pretty different systems—n8n is all about no-code, user-friendly automation, while frameworks like LangGraph give you much more granular control if you’re comfortable with programming26. If you need deep agent-to-agent logic and custom state handling, code frameworks will always do that better, but n8n keeps getting better and with smart prompt design, you can actually pull off more advanced handoffs than most people realize
A multi agent system that shares information like a swarm system is possible with n8n.
In n8n, you can set up two agents that are both using the same memory. I prefer using a chat history table in supabase.
If both or all agents share memory, they will act very much like the swarm system.
If this response helped you, please click the heart to show that it is useful If this response solved your issue, mark it as the solution to help the community
I think he is referring to the fact where at least two nodes are linked between them in “full mesh” mode, so there’s a proper hand off support for that.
Yeah, I’m not sure how much state sharing I really need. Plus having multiple agents with large and very different system prompts usually makes less context / more relevant context better rather then just having all the previous memory thrown into every chat conversation. I would love it if I could orchestrate a hand-off to an agent with a fresh state without having to leave the chat window. Almost like being able to navigate between different Claude projects (each with their own tools and system prompt] using natural language. That way when my clients come into work to use my systems they can just type “lets write a blog” or “lets go through emails” and interact with an agent with a fresh memory and no intermediary agent translating subagents like tool calls.
If this response helped you, please click the heart to show that it is useful
If this response solved your issue, mark it as the solution to help the community