GroupChatOrchestration uses too many tokens when MaximumInvocationCount > 3

[alliscode]

Can guidance on the MaximumInvocationCount = ? for multi agents for demos
We have found going above 3 we utilise so many tokens...
Do you have any guidance?

GroupChatOrchestration orchestration =
            new(
                new RoundRobinGroupChatManager()
                {
                    MaximumInvocationCount = 3
                },
                evangelistAgent,
                contentReviewerAgent,
                publisherAgent
            )
            {
                //LoggerFactory =  LoggerFactory.Create(builder => builder.AddConsole().SetMinimumLevel(LogLevel.Warning)),
                ResponseCallback = monitor.ResponseCallback,
                StreamingResponseCallback = monitor.StreamingResultCallback,
            };

ALso can you please confirm in the thread output window what is the length? what do this represent?

[crickman]

In SK, the sample was updated to compensate for the fact that the RoundRobinGroupChatManager will always run to MaximumInvocationCount without any other termination logic. This issue exists in Python also.
roupChat/RoundRobinGroupChatManager.cs

https://github.com/microsoft/semantic-kernel/blob/main/dotnet/samples/GettingStartedWithAgents/Orchestration/Step03_GroupChat.cs

The root cause is that RoundRobinGroupChatManager has a design bug in that it isn't usable as is (perhaps should be abstract?).

This is the same issue in agent-framework (both Python and .NET):

https://github.com/microsoft/agent-framework/blob/main/dotnet/src/Microsoft.Agents.Orchestration/GroupChat/GroupChatManager.cs
https://github.com/microsoft/agent-framework/blob/main/dotnet/src/Microsoft.Agents.Orchestration/G

[TaoChenOSU]

The default RoundRobinGroupChatManager is designed to run with a predefined number of rounds. If users need to ability to terminate early, they should extend the RoundRobinGroupChatManager and implement a custom termination strategy by overriding the should_terminate method in Python (ShouldTerminate in .Net).

[TaoChenOSU]

More information on "too many tokens" would be helpful. Does the customer see a sudden jump when they go from 2 to 3 max rounds?

[kinfey]

Tao Chen (@TaoChenOSU) 2 question

1. Is GroupChatOrchestration still available in Python? https://github.com/microsoft/semantic-kernel/blob/main/python/samples/getting_started_with_agents/multi_agent_orchestration/step3_group_chat.py

2. While using RoundRobinGroupChatManager in .NET, I discovered that if the agent is CodeInterpreterTool, how can I obtain the files generated by the agent? I'm currently unable to do so.

02-dotnet-agent-framework-multiagents-groupchat.ipynb

If you use Foundry's SDK, can you get the generated ，like this

02-dotnet-agent-framework-multiagents-groupchat-testing.ipynb

[joslat]

Imho, adding the GroupChat - GroupChatOrchestration we are again implementing things in code and not with a workflow object (I mean inside the groupchat what happens is akin to a black box as it is not related to the workflow, right? - we will see what happens with the concrete agants but that's it).

I am pretty confident that we will not see the internals of the GroupChat in DevUI and the .ToMermaidString does not show the flow within the GroupChat.

I've implemented this pattern with proper workflow and I believe this should be the way to go - this or we find a way to expose the internals of the groupchat to components like DevUI and worfklow graph renderers, so we can follow them "inside".

Something like:

See #1790