Agents in AgentGroupChat not applying each others feedback

[mickbut-ler]

Hey everyone, I’m running into an issue with AgentGroupChat where my two agents; ChefAgent and ReviewerAgent, aren’t actually feeding each other’s feedback.

1. In my ReviewerInstructions I’ve added MAKE SURE THE RESULT HAS AT LEAST ONE FRENCH WORD; otherwise return isValid false and include errors. ReviewerAgent correctly returns isValid=false and populates errors when there’s no French word, but ChefAgent (which runs next) never applies that feedback, so the chat gets stuck in a loop. Am I missing something to include the feedback form one agent to another?

2. My termination strategy is supposed to stop after 4 iterations, but the conversation continues indefinitely;

Here’s the core of my setup:

private static AgentGroupChat SetupGroupChat(Kernel kernel)
{
    const string ChefAgentName = "ChefAgent";
    var chefAgent = CreateAgent(
        ChefAgentName,
        ChefInstuctions,
        kernel.Clone(),
        typeof(IngredientsData)
    );

    const string ReviewerAgentName = "ReviewerAgent";
    var reviewerAgent = CreateAgent(
        ReviewerAgentName,
        ReviewerInstructions,
        kernel.Clone(),
        typeof(ValidationResultDto)
    );

    // Alternate turns: Chef → Reviewer → Chef → …
    var selectionFunction = AgentGroupChat.CreatePromptFunctionForStrategy(
        """
        Determine which participant takes the next turn based on the most recent speaker.
        Only respond with the agent name. No one speaks twice in a row.

        Participants:
        - ChefAgent
        - ReviewerAgent

        Rules:
        - If no prior turn, ChefAgent goes first.
        - After ChefAgent, ReviewerAgent goes next.
        - After ReviewerAgent, ChefAgent goes next.

        History:
        {{$history}}
        """,
        safeParameterNames: "history"
    );

    var selectionStrategy = new KernelFunctionSelectionStrategy(selectionFunction, kernel)
    {
        InitialAgent = chefAgent,
        ResultParser = result => result.GetValue<string>() ?? ChefAgentName,
        HistoryVariableName = "history"
    };

    // Terminate when ReviewerAgent responds with “IS_DONE”
    const string TerminationToken = "IS_DONE";
    var terminationFunction = AgentGroupChat.CreatePromptFunctionForStrategy(
        $$$"""
        If the last ReviewerAgent response contains "{{{TerminationToken}}}", reply exactly with "{{{TerminationToken}}}".
        """,
        safeParameterNames: "history"
    );

    var terminationStrategy = new KernelFunctionTerminationStrategy(terminationFunction, kernel)
    {
        Agents = [ reviewerAgent ],
        HistoryVariableName = "history",
        MaximumIterations = 4,
        ResultParser = result => result.GetValue<string>()?.Contains(TerminationToken, StringComparison.OrdinalIgnoreCase) ?? false
    };

    return new AgentGroupChat(chefAgent, reviewerAgent)
    {
        ExecutionSettings = new() 
        {
            SelectionStrategy = selectionStrategy,
            TerminationStrategy = terminationStrategy
        }
    };
}

[sophialagerkranspandey]

Thanks for providing this information. There are a couple of questions that will help with our troubleshooting. What model are you using and are you using AzureOpenAI or OpenAI?