Node attributes vs state

[ag14774]

Question

Could you explain the reasoning or guidelines for deciding when to use an attribute within a node versus using a mutable state? For instance, in the genai_email_feedback.py example provided in the documentation, both approaches are utilized, but the criteria for choosing one over the other are unclear.

Additional Context

No response

[DouweM]

@ag14774 The State object is passed to every node, so you can use it for data known upfront that different nodes may need access to (like the User in the example), or for keeping track of things that happened during the agent run that different future nodes may need access to (like the message history). The State can also be accessed after the agent run.

Attributes on nodes store the function arguments that are passed to the node when it's instantiated (e.g. WriteEmail(email_feedback=result.output.feedback)). They are used to communicate from one node to the next exactly what it should do -- like function arguments elsewhere. They are not available to any other nodes (unless they're explicitly passed through).

So agent state is for stuff that's relevant to the entire agent run, and node arguments are for stuff that's only relevant to the specific node being instantiated.

[ag14774]

Thank you for the reply.

If the User is known in advance, and there is no plan to mutate it, why not just use it as a dependency instead?

In addition, in the first example where the number is incremented until it is divisible by 5, it really feels like the graph has a state even though it doesn't. In some sense, foo seems to be the state, it 'feels' like a state, and it could be implemented as a state as well (it's also global to the whole graph even).

For what it's worth, a small paragraph clarifying when to use a state, when to use attributes, and when to use dependencies, would be nice.