Python: Bug: AgentGroupChat | ChatCompletionAgent | Complete chat history is getting passed to llm

[pintuiitbhi]

Describe the bug
There is two agent - Agent1 and Agent2. I am using AgentGroupChat using ChatCompletionAgent. It is multi turn conversation based. Then complete chat history is getting passed to llm. Each agent has its own Kernel instance.
This is what currently I observerd from two DEBUG logs of Semantic-Kernel

This is the each item content in json_data for a llm api calls request has -

ROLE, CONTENT

1. system, Agent1 Prompt
2. user, task
3. assistant, tool_calls
4. tool, tool_calls result
5. assistant, Agent1 response
6. assistant, Agent1 response
7. assistant, Agent2 Prompt
8. assistant, Agent1 response
9. assistant, Agent1 response
10. assistant, Agent2 response

As its clear for more iterations it will pass more such messages and we will get token limit exceeded error.
Currently for my use case I just need this for any llm api call

1. system prompt
   2, user task
2. assistant tool_calls
3. tool response
4. assistant response
5. assistnat response

To Reproduce
Steps to reproduce the behavior:

1. This can be observed in Semantic Kernel DEBUG log for any multi-turn group chat for atleast 10 iterations

Expected behavior
6 message items should be enough.

When its turn then this should be the llm api call payload:

1. system prompt
   2, user task
2. assistant tool_calls
3. tool response
4. assistant response
5. assistnat response

When its turn then this should be the llm api call payload:

1. system prompt
   2, user task
2. assistant tool_calls
3. tool response
4. assistant response
5. assistnat response

Screenshots
If applicable, add screenshots to help explain your problem.

Platform

• Language: Python
• Source: main branch , 1.30.0 Semantic kernel version
• AI model: Currently I used gpt-4o using Azure
• IDE: Vscode
• OS: Windows

Additional context
Add any other context about the problem here.

[moonbox3]

Hello @pintuiitbhi, please have a look at our new group chat orchestration patterns (step3): https://github.com/microsoft/semantic-kernel/tree/main/python/samples/getting_started_with_agents/multi_agent_orchestration

We have done a soft-release of these patterns and plan to have a more structured blog post/migration guide coming soon. We won’t be putting any more dev effort into the old AgentGroupChat patterns. We recommend devs switching to the new patterns as soon as possible. We will be tracking feedback on these new patterns going forward.

[moonbox3]

As a quick note: the "old/legacy style" AgentGroupChat does pass the entire chat history to the model. That is by design. As I mentioned above, we're moving to the new orchestration patterns. Please have a look at those, and if you need help create a new issue for us to track. Closing this issue as there won't be more development on the old patterns.

[pintuiitbhi]

I have migrated now to group chat orchestration and created a custom chat manager inheriting from group chat manager following one of sample example.

I still see the same issue. SK is passing complete chat history to llm api chat completion api.