Using a saver across multiple agents

[cwoolum]

When using the supervisor/agent pattern, I'm wondering what's the best way to share a checkpointer across multiple agents? I'd like all of my agents to share the same supervisor thread id so we can trace the flow across all agents within a graph. I've tried explicitly passing configurable to the child invoke but I don't think that's doing what I intend. Ideally, I'd like the supervisor checkpoint_id to be the parent_checkpoint_id of the agent.

[vbarda]

@cwoolum if the agents are a part of the same parent graph, it should be already propagated and traced correctly. parent_checkpoint_id serves a different purpose - it's the checkpoint ID of the previous checkpoint, so i don't think it's what you want in this case

are you running into any specific tracing issues with this?

[cwoolum]

I'm not sure if I'm running into issues related to that yet, I'm still debugging on my end. When calling a child agent, should the parent configurable always be passed into the child invoke?

[vbarda]

hm, are you manually calling the child agent? usually child agents are added as subgraphs, and the config is automatically propagated to all nodes in the graph (including subgraphs). and then those subgraphs are just invoked the same way as any other node (by the library)

[cwoolum]

Yes, I manually call the child agents because otherwise the entire supervisor context gets passed to the children. I detail the approach here

#1008 (comment)

This lets me optimize context sizes. I think this fixes it though

    async (state: SupervisorState, event) => {
      const inputMessages =
        await childAgentPromptTemplate.formatMessages({
          someRequiredProp: JSON.stringify(state.somethingThatNeedMapped, null, 2),          
        });

      const output = await childAgent.invoke({
        messages: inputMessages,
      },
      {
        // Pass it through verbatim like the supervisor library
        configurable: event.configurable,
      });
      let { messages } = output;

      messages = messages.slice(-1);

      messages.push(
        ...createHandoffBackMessages(childAgent.name!, 'supervisor'),
      );

      return new Command({
        update: {
          childAgentOutput: output.structuredResponse,
          messages: messages,
        },
      });
    },
    {
      subgraphs: [childAgent],
    },
 

[vbarda]

even if you call the agent manually like in your example above, it's still a node function, and the config is automatically propagated to all nodes and all subgraph invocations inside nodes. you shouldn't need to pass config to childAgent.invoke. at least that is the case in Python, but perhaps there are some differences in JS cc @benjamincburns

[benjamincburns]

you shouldn't need to pass config to childAgent.invoke. at least that is the case in Python, but perhaps there are some differences in JS

@vbarda @cwoolum unfortunately in JS it depends. If you're running in nodejs, this is correct - you don't need to forward the config object. It'll get propagated behind the scenes via the async local storage facility from node's async_hooks module. Unfortunately, ALS isn't widely supported outside of node and node-compatible runtimes like deno. In the browser or in react native, you'll need to pass the config yourself.

There's an open feature request in #879 for this, but so far we haven't found a viable solution, outside of the async-context tc39 proposal.