Asynchronous chatbot flow with AgentExecutor and PlanAndExecute?

[Sam1320]

I have a bot which I am running serverless. It uses tools to answer queries BUT it needs to stop at every step and ask the user to verify if its next action makes sense.

What I did was see how the _call() function of langchain.agents.agent.AgentExecutor was implemented. Based on that I used a LLMSingleActionAgent that outputs (Action, Observation) tuples which I then store on a list which is then updated and passed again to the agent as history on the next iteration once the user provides feedback (which can be e.g. 1 day later).

I now want to expand and generalize more my bot to behave more like an autonomous general task solving agent and follow either the babyagi or autogpt approach. However now I find myself again scratching my head going into the rabbit-hole of the the _call implementation of langchain.experimental.plan_and_execute.PlanAndExecute from Plan and Execute but this time its one deeper level of nestendess since each Step requires a full AgentExecutor loop.

I am starting to "unwrap" the code to see how to best modularize it but its a lot of work to try to follow all the layers and I can't help but think that there has to be a better way to do this or that I might be missing an already implemented solution for this common use case.

So:

1. Is there an already implemented solution either in LangChain or somewhere else using LangChain primitives?
2. If answer to 1. is no: Is my approach mentioned above good?, Is there a simpler, more modular way to implement this than trying to reverse engineer the _call() functions of the Executor Chains?

Thank you very much in advance!

[batmanscode]

Maybe this can help:

https://python.langchain.com/docs/modules/agents/tools/how_to/human_approval