ReACT Agents and Custom States

[sharyar]

This question is less about what is possible but more about what is the right way to do things and make sure they work.
As I read the docs and examples such as the "Customer Support Agent" I see that most react style tool calling agents are implemented with a fairly simple AgentState with messages as an attribute.

However, I also see that tools can return artifacts and also receive and update state (via injected args and commands).

I am currently trying to build a ReACT style agent that has a bunch of tools at its disposal. The problem/confusion I am running into is how to design my tools and states based on this criteria:

1. The tools are often used one after another, however some tool outputs may need feedback prior to continuing to the next tool.
2. The tools are often complex and may return nested pydantic model instances and/or utilize pydantic model instance in their args. Here is an example:

@tool
def cool_tool(arg1: PydanticModel_A, arg2: PydanticModel_B) -> PydanticModel_C: 
    ...

Now, I am hoping that the model will take the results it gets from one tool and pass them to the other without trying to "generate" them on the fly. It sometimes does but sometimes not. In order to enforce this behaviour, I thought: why not update the AgentState from within the tool call by using Command as the return. This also means that the AgentState needs to have new attributes that can store these results. However, this feels off to me.
And then, I notice the llm will not take these new attributes into account and would instead try to generate the args for the subsequent tool calls based on the "ToolMessage" that it had recieved in the messages list. It seems that the llm only sees the "messages" and is not aware of the rest of the state.

So. My question is:

1. What is the right way to build a ReACT style agent where the tool calls can generate fairly complex outputs in the form of PydanticModelInstances? Does it make sense to have custom state attributes for ReACT style agents or is that more suitable for workflows?