How to combine multipe Agents in one Chat

[notdefine]

Hi all,

I used the good tutorial Video and the Docs, but I didn't found a recommendation howto combine multiple Agents in one Chat. I need this for a case, in which a customer enter a message and the Agent should decide if it is a task for him. The Agents habe absolute different context, so one general Agent with multiple Tools should not be the right way.

I want to do somethind like this (pseudocode)

$msg = new UserMessage($message);

$orderAgent = OrderAgent::make();
if ($orderAgent->couldHandle($msg)) {
    return $orderAgent->chat($msg)
}

$youTubeAgent = YouTubeAgent::make();
if ($youTubeAgent->couldHandle($msg)) {
    return $youTubeAgent->chat($msg)
}

I build one hacky solution, I instruct the agents to Response with "Not my Task" if there is nothing to do for them and running each Agent with this. But maybe there is a better solution.

Do have any hints for me? Thanks

[ilvalerione]

I think this is where Workflow should comes into play: https://docs.neuron-ai.dev/workflow/getting-started

It could not be easy at the beginning, but once you start experimenting with it you will also able to follow tutorials from other ecosystems like LangGraph and implement your own multi agent system with easy in PHP.

[notdefine]

Okay, my Problem was, that in DetermineMealNode in the run() method the Agents wer initialized. So every new User Text Input the start without the history and so they forget the last input. I correct the example in my GIT Repo.

[ilvalerione]

To manage the chat history at the workflow level you should use the chat history as a stand alone component:

$state = new WorkflowState();
$state->set('chat_history', new FileChatHistory(__DIR__, 'workflow'));

// Run the workflow passing the initial state
$workflow = new SimpleWorkflow();
$state = $workflow->run($state);

Inside nodes you can use the chat history methods to add and retrieve messages:

class ExampleNode extends Node
{
    public function run(WorkflowState $state): WorkflowState
    {
        $message = new UserMessage(...);
        
        $state->get('chat_history')->addMessage($message);
        
        $response = $agent->chat($message);
        
        $state->get('chat_history')->addMessage($response);
        
        return $state;
    }
}

The internal agents like MealOrderAgent just simplify the interaction with AI using their built-in features: tool calls, structured output, etc.

[ilvalerione]

In the latest version of NeuronAI "1.16.13" the persistence layer uses serialization to store and retrieve the workflow state during interruption. That means that if you use the Chat History component as a state field it will be deserialized as a Chat History instance, so you can continue to add and retrieve messages.

I thought this wass something required for the framework. Let me know if it also helps in your scenario.

[notdefine]

Yes, with the serialization the history objects get keeped. Thanks

[thomasdv]

@ilvalerione that however broke SQLChatHistory as it is now trying to serialize PDO.

edit: but there is no reason to keep that one in the $state, so no problem there, sorry :)

[bilogic]

Interesting, are these RAG agents? Why is there a need to separate them into 2 agents?