Issue with ApprovalRequiredAIFunction in sub-agent

[hobbsax1]

I am using ApprovalRequiredAIFunction to try and perform human in the loop actions.

I have the following setup:

TriageAgent
UserRoleAgent

TriageAgent using UserRoleAgent as a tool.

UserRoleAgent has several tools, one of which is AddUserToRole that is encapsulated in ApprovalRequiredAIFunction

When I try to get the approval, the TriageAgent seems to be eating the FunctionApprovalRequestContent content coming back. I have middleware on both agents and can see where the UserRoleAgent has the last message as a FunctionApprovalRequestContent with the proper argument calls, but I cannot seem to bubble that up so that the triage agent returns it to the caller.

We are using a custom ASP.NET Core with restful endpoints so we have an endpoint /agent that takes (approved takes an optional approved status if the user was asked to approve a tool call):

public class Request 
{
    public string Message { get; set; }

    public bool? Approved { get; set; }
}

And then responds with either the messages from chat or info needed for the approval (the message is generated from the function call arguments and name in the backend):

public class Response
{
    public string? Message { get; set; }

    public bool ApprovalRequired { get; set; }

    public string? ApprovalMessage { get; set; }
}

Anyone know of a way to bubble up approval requests from sub agents that are acting as tools themselves?

[moonbox3]

Adding @markwallace-microsoft to help route appropriately.

[alsetr13]

Hi, were you able to figure out how to bubble up approval properly ?

[hobbsax1]

Yes, in the prerelease versions the ThreadChat had an append method that I use to rebuild the proper user/system/assistant message flow (I take messages from the sub chat and add them to the top level chat such that it doesn't 400 going to OpenAI). The newer version switched away from the Append to a Set (make it so) on the thread so I just had to take the existing messages and add the ones I wanted to "append" and then set all of them back. Ended up being simple fix but took a long time troubleshooting.

[ishita-0301]

I haven't tried this exact setup, but my instinct would be to treat approval requests as first class events instead of normal chat messages. If the sub-agent emits a "FunctionApprovalRequestContent", I'd propagate that state all the way to the top level and pause execution until the user responds, rather than expecting the parent agent to interpret it correctly.

This kind of propagation is one of those runtime edge cases that's easy to miss in multi-agent systems. I've been looking at https://github.com/FailproofAI/failproofai for similar ideas around handling execution failures and interruptions in a more structured way.