LLM calling node name of agent graph create by prebuilt.create_react_agent

[ghost]

Hello, this is Yoo.
I've been used prebuilt.create_react_agent to make agents and connected them to build multi-agent systems.

It was super cool, but the thing happened when I tried to stream a result of one of the agents (which was supervisor agent in my case). I wrote a streaming code with stream_mode="messages" to stream tokens as they are produced by the LLM. But below is the result of chunk & metadata which is the output of streaming

content='\n' additional_kwargs={} response_metadata={} id='run--f567f474-6ee8-493f-b533-0a89269187a7'
{'thread_id': '975250e8-36a9-47af-afc5-0fb7420fa0db', 'langgraph_step': 3, 'langgraph_node': 'agent', 'langgraph_triggers': ('branch:to:agent',), 'langgraph_path': ('__pregel_pull', 'agent'), 'langgraph_checkpoint_ns': 'research_agent:f9beb20b-4112-7f3e-227a-5392c6a11448|agent:1bc3291f-4200-e125-bc8f-906557617fa0', 'checkpoint_ns': 'research_agent:f9beb20b-4112-7f3e-227a-5392c6a11448', 'ls_provider': 'openai', 'ls_model_name': 'gpt-4o-mini', 'ls_model_type': 'chat', 'ls_temperature': 0.5}
content='-' additional_kwargs={} response_metadata={} id='run--f567f474-6ee8-493f-b533-0a89269187a7'
{'thread_id': '975250e8-36a9-47af-afc5-0fb7420fa0db', 'langgraph_step': 3, 'langgraph_node': 'agent', 'langgraph_triggers': ('branch:to:agent',), 'langgraph_path': ('__pregel_pull', 'agent'), 'langgraph_checkpoint_ns': 'research_agent:f9beb20b-4112-7f3e-227a-5392c6a11448|agent:1bc3291f-4200-e125-bc8f-906557617fa0', 'checkpoint_ns': 'research_agent:f9beb20b-4112-7f3e-227a-5392c6a11448', 'ls_provider': 'openai', 'ls_model_name': 'gpt-4o-mini', 'ls_model_type': 'chat', 'ls_temperature': 0.5}

The langgraph_node of metadata of all agents are the same, which is 'agent'. This is because in the code of create_react_agent', the 'agent is hardcoded as the name of llm calling node. So I was not able to distinguish the chunk output of each agent by node name.

By the way, there is name parameter which is the name of graph of agent. But it does not ensure in this kind of case. Since I'm not familiar with langgraph enough, there can be the reason why the code is written this way. And also there might be some better ways to LLM node of each agent's graph. That's why I need maintainer's advice.

If this makes sense and it's judge that new feature is needed for this case,
I really hope to get a chance to request a PR and receive some feedbacks! 😃

Best Regrads,

[Icyrockton]

Same, I want to track the output of all nodes. It needs a parameter to be able to change the default node name.

node_name = run.metadata.get('langgraph_node', '')