I find LangGraph too rigid for our Agentic B2B Product

[teshnizi]

We're building an AI b2b saas product.

For our agentic systems, I find langgraph too rigid to build enterprise apps on. For instance, we would like to customize agents and handoffs for each customer, but langgraph expects you to hardcode handoffs (e.g. transfer_to_agent_a() rather than transfer_to_agent(agent_a)).

This means we will have file level customization (acme_inc.py) rather than configuration level customizations (acme_inc.json), which does not sound like a good practice.

Has anyone else experienced that too or am I thinking about it wrong?

[hinthornw]

Appreciate the feedback! You can jump to any node in the graph by returning a command from a node: Command(goto="foo")

https://langchain-ai.github.io/langgraph/how-tos/command/?h=command

The world is your oyster.

Also totally valid to define deep nodes if you want. LangGraph nodes define snapshot boundaries for things like human-in-the-loop but you don't have to use a thousand nodes in a graph if you don't want.

[teshnizi]

Thanks!
So I've seen that but I want the agent to make the handoff decision. I guess I can use a structured output call in the agent call to get the next tool, but I'll have to create an output structure for all the invoke calls. That doesn't sound like a clean solution. Am I missing anything?

[hinthornw]

It sounds like you may be wanting a high-level API that also gives you control over handoffs? You could try something like langgraph-swarm https://github.com/langchain-ai/langgraph-swarm-py as well.

You could also write an llm wrapper to go from structured output -> command in a few lines of code if you want to make your own higher-level syntax sugar perhaps?

[teshnizi]

Yeah I think a generalized create handoff tool from swarm is what I needed. I think I figured out how to implement it based on the your previous reply. Thanks!

[vbarda]

@teshnizi there is nothing in langgraph that defines the idea of "transfer_to_agent_X" - langgraph is a low level library where you can implement any possible variants of agent architectures / handoffs.

you can almost certainly implement a general purpose handoff tool that has agent name populated by an LLM, if that's what you're looking for

[teshnizi]

Yeah that's what I'm looking for. Do you have any thoughts on whether each agent should have its own tool node or all of them should use the same tool node?

[vbarda]

I think usually you want each agent to have a separate tool node, and encapsulate each agent in a separate subgraph