How to include human in the loop AgentGroup of Agent Framework of Semantic Kerenel

[MahorShekhar]

We are using AgentGroupChat feature of Agent framework. While running a small POC, we realised it doesn't include human in the loop while solving assigned task. It just takes a user prompt in the beginning and jump on to addressing that query using multiple agents. What we need is a back and forth human interaction with the agents where user can provide his input at each step or agents can ask clarification questions.

Any help or suggestion is very much appreciated.

Thanks.

[moonbox3]

We are working on a user proxy agent for our new agent group chat patterns that should be available in the coming weeks.

[MahorShekhar]

@moonbox3 That sounds exciting! We're eager to integrate the new agent group chat patterns into our work. Do you have an estimated timeline for their release? This would help us plan accordingly, as it's a key dependency for us.

[moonbox3]

We've done a soft-launch of some new orchestration patterns for Python and .NET:

• Python: https://github.com/microsoft/semantic-kernel/tree/main/python/samples/getting_started_with_agents/multi_agent_orchestration
• .NET: https://github.com/microsoft/semantic-kernel/tree/main/dotnet/samples/GettingStartedWithAgents/Orchestration

Each have a sample of using a human in the loop for the group chat. New docs for these patterns will be released soon.

[jupiver123]

@moonbox3, InteractiveCallback though allows user to input additional information, however we find it is hard to use in a web application environment.

In our setup, a UI is a web application where user will post texts to an API endpoint. User prompt will be passed to an orchestrator (we are experimenting HandoffOrchestration and GroupChatOrchestration) to process. During that, it may need additional information from the user (e.g., to further clarify ambiguous keyword). It is naturally to pause the conversation and return to web app so user can enter those refinements. In this, it is hard to use InteractiveCallback.

For now, we instantiate the orchestration to take ChatHistory as an input (e.g., to use 'new HandoffOrchestration<ChatHistory, string>(...)') with the hope that the orchestration can retain the ongoing conversation. Can you comment about this workaround? or suggest an alternative. Thanks.

[moonbox3]

I will add @crickman for further help on the .NET side.

[crickman]

@jupiver123 - Other customers have registered the same ask. We're actively looking at an improvement here. Hope to have a direction defined over the next couple days.