Customer Support Bot - How to add confirmation from customer, not agent inside (which like admin role) by interrupt_before in LangGraph

[weiminghuaa]

I reviewed the customer-support.ipynb sample, and it closely aligns with what I need. It seems to be a powerful tool for building an AI customer agent for my business.

However, I noticed that in the sample, all confirmations are handled within the agent itself. Some operations in my use case require explicit customer confirmation, but based on the guide, it appears that only the agent (or an admin) can approve these actions, rather than delegating the confirmation back to the customer.

For example, some operations—such as transferring money to a friend—require explicit user confirmation before being executed. Ideally, the agent should generate a bill or transaction request, prompt the user for confirmation, and only proceed once the user approves it.

ps: I have an idea to design the process in agent flow, instead of using interrupt_before in LangGraph.

[weiminghuaa]

After many attempts, I find it's also very hard to add confirmation in agent flow or agent chain.

User input: transfer request like "Transfer to Lucy 100 for happy birthday"
->
Bill generation node: create a bill from the input. (many things need to identify, user, amount, note, etc. need human loop-in while missing any mandatory data)
->
Human loop-in: Ask user to confirm the bill
->
Execute bill node: execute the bill transaction
->
Human loop-in: If it'sfailed need ask user to retry, maybe

I failed to set up a flow or chain like that, hope someone could give me some guidance.

[heiyuxiaokai]

@weiminghuaa I want to implement similar workflows. Can you share how to build?

[weiminghuaa]

the document said you can create a api to ask customer approve and make the interrupt continue. another way is to abandon the use of interrupt from langchain, set up intructions in your prompt to implement all flows. i choose the later one.

[heiyuxiaokai]

Thanks for your sharing. I am trying the interrupt-based approach as
https://github.com/langchain-ai/langgraph/blob/main/docs/docs/tutorials/get-started/4-human-in-the-loop.md
But I am confused how to achieve these just through prompt. In this way, humans' questions are new to the chatbot even with chat history. Can you share a simple example?

[weiminghuaa]

sorry i can't give you an example, i remembered it was just follow the real journey to design prompt. don't spend much time on it if it's not useful for you. Good luck!