Python: Challenges with AgentGroupChat Post-Upgrade

[vikeshsingh37]

We are working on a multi-agent system using AgentGroupChat from Semantic Kernel. Our agents were designed to receive all chat history, which worked fine with version 1.20.0. However, after upgrading to 1.27.1, our system started breaking. We are using the AgentGroupChat.invoke function that internally uses ChatHistoryChannel. We noticed a change in chat_history_channel.py. Previously, it sent the entire chat history, but now it only sends the last message, causing errors in our code. Before making any changes, we wanted to explore possible options with you. Could you please let us know if there is another mechanism to provide the whole chat history to the agent?

[moonbox3]

We're working on new abstractions/patterns for group chat (which are still in the experimental phase, thus changes may be introduced). Please stay tuned as they should be available soon.

[sushaanttb]

We're working on new abstractions/patterns for group chat (which are still in the experimental phase, thus changes may be introduced). Please stay tuned as they should be available soon.

@moonbox3
As we are blocked because of this new change, we would like to understand how can we get the entire chat history? as unlike before, now in the new version of SK, AgentGroupChat is only passing the last message from ChatHistory.

We would like to understand the rationale behind this and if this is expected to be the contract going forward.

Are we expected to use the thread object here? If yes, what are the guidelines as the documentation doesn't talks about thread concept from the AgentGroupChat level and also, it would be a good idea to also call out these new changes in our documentation.

[moonbox3]

I'm looking into this further. The reason why we're now sending messages=self.messages[-1] is because the context of the conversation is already contained in the thread.

In an AgentGroupChat example between an AzureAIAgent and a ChatCompletionAgent, I see that the ChatCompletionAgent is getting the full context.

Here's an example output I see on the console (in the middle of a conversation between two agents):

# AuthorRole.USER: 'a slogan for a new line of electric cars.'
# AuthorRole.ASSISTANT - CopyWriter: '"Drive Smarter, Not Harder."'
# AuthorRole.ASSISTANT - ArtDirector: 'This headline has punch and brevity, but it leans on a familiar cliché without clearly communicating a benefit or unique selling proposition. Refine it by specifying what makes driving smarter — is it efficiency, technology, savings? Sharpen the copy to reflect a concrete advantage or differentiator, and ensure the message aligns closely with the brand’s promise.'
# AuthorRole.ASSISTANT - CopyWriter: '"Pulse of the Future: Where Power Meets Precision."'
# AuthorRole.ASSISTANT - ArtDirector: 'This line has energy and a forward-looking attitude, but it relies on vague abstractions—“pulse of the future,” “power,” “precision”—that could apply to any number of products or industries. Refine by grounding the message in a tangible benefit, unique feature, or emotional payoff tied directly to the offering. Bring clarity on what “power” and “precision” mean for the customer.'
# AuthorRole.ASSISTANT - CopyWriter: '"Charge Forward: Innovation at Every Turn."'

If I examine the ChatHistory we're using as part of ChatCompletionAgent invocation (being built from the thread), I see:

There are 5 messages included in the current ChatCompletionAgent invoke method, some messages from the AzureAIAgent, and some messages for the ChatCompletionAgent. This is expected and how it was handled previously - before introducing the AgentThread concept.

Can you please let us know what is blocking here? Thanks.

[moonbox3]

I do see our documentation is out of date. We can get that updated. Thanks for pointing that out.

[sushaanttb]

I'm looking into this further. The reason why we're now sending messages=self.messages[-1] is because the context of the conversation is already contained in the thread.

Okay, and yes that's what I also eventually inferred but wanted to confirm.

In an AgentGroupChat example between an AzureAIAgent and a ChatCompletionAgent, I see that the ChatCompletionAgent is getting the full context.

Here's an example output I see on the console (in the middle of a conversation between two agents):

# AuthorRole.USER: 'a slogan for a new line of electric cars.'
# AuthorRole.ASSISTANT - CopyWriter: '"Drive Smarter, Not Harder."'
# AuthorRole.ASSISTANT - ArtDirector: 'This headline has punch and brevity, but it leans on a familiar cliché without clearly communicating a benefit or unique selling proposition. Refine it by specifying what makes driving smarter — is it efficiency, technology, savings? Sharpen the copy to reflect a concrete advantage or differentiator, and ensure the message aligns closely with the brand’s promise.'
# AuthorRole.ASSISTANT - CopyWriter: '"Pulse of the Future: Where Power Meets Precision."'
# AuthorRole.ASSISTANT - ArtDirector: 'This line has energy and a forward-looking attitude, but it relies on vague abstractions—“pulse of the future,” “power,” “precision”—that could apply to any number of products or industries. Refine by grounding the message in a tangible benefit, unique feature, or emotional payoff tied directly to the offering. Bring clarity on what “power” and “precision” mean for the customer.'
# AuthorRole.ASSISTANT - CopyWriter: '"Charge Forward: Innovation at Every Turn."'

If I examine the ChatHistory we're using as part of ChatCompletionAgent invocation (being built from the thread), I see:

There are 5 messages included in the current ChatCompletionAgent invoke method, some messages from the AzureAIAgent, and some messages for the ChatCompletionAgent. This is expected and how it was handled previously - before introducing the AgentThread concept.

Thanks & this is essentially because the chat history object is getting updated from thread first (line no. 326 from your shared screenshot) because the thread object is being maintained by Agents internally which is leveraged accordingly to build the chat history.

Can you please let us know what is blocking here? Thanks.

Actually we had created a custom implementation of our Agent class where we had overridden the invoke() method
and this was working fine with previous SK version but post our upgrade to new SK version we noticed this breaking change
as we were not making use of thread object.

[moonbox3]

Thanks for your response, @sushaanttb.

Please keep an eye on which modules and features are marked as experimental versus those tagged as release_candidate. Features typically start as experimental, and as they mature toward general availability (GA), they're promoted to release_candidate.

We also call this out on our Learn site:experimental features carry a note indicating their APIs may change significantly as we work toward GA.

There was a progression of this for our single-agent abstractions between SK 1.20.0 and the GA release in 1.27.0.

Do note that our Agent Group Chat patterns are still experimental. It's only once something reaches GA that we consider the APIs stable, with no expected breaking changes.

We understand developers want to build on stable APIs, and we do our best to avoid disruptions. That said, some breaking changes are necessary in the journey to GA so we can ensure long-term stability.

Is there any further action required on this issue or can we close it?

[munishm]

Good to close this issue. Thank you