Make a triage agent to make a decision which agent to call without using LLM in AutoGen 0.4

[rubans]

I wanted to create a simple example of multi agents where a planning agent (agent1) decides which to agent to call next (agent 2 or agent 3). All 3 agents for this purpose would be user agents that will not use LLMs, so not using Assistant Agent.

I have looked at Handoff patterns and Swarm Teams but not sure really how to do it, they all seem to rely on Assistant Agent with LLM making the decision.

For e.g.
We have 2 agents - Agent 2 which will count down from 10 and Agent 3 which will count up to 10.
Agent 1 takes a user input of to decide which will decide which option user wants to do i.e. count down or count up

From my limited knowledge of autogen so far, I have been building custom agents based on BaseChatAgent:

class CountDownAgent(BaseChatAgent):
    def __init__(self, name: str):
        super().__init__(name, "A simple agent that counts down.")

    @property
    def produced_message_types(self) -> Sequence[type[ChatMessage]]:
        return (TextMessage,)

    async def on_messages(self, messages: Sequence[ChatMessage], cancellation_token: CancellationToken) -> Response:
        if not messages:
            return Response(chat_message=TextMessage(content="No number received!", source=self.name))
        last_message = messages[-1].content.strip()
        if not last_message.isdigit():
            return Response(chat_message=TextMessage(content="Please enter a valid number.", source=self.name))
        count = int(last_message)
        return await self.countdown(count)

    async def countdown(self, count: int) -> Response:
        inner_messages: List[ChatMessage] = []
        for i in range(count, 0, -1):
            msg = TextMessage(content=f"{i}...", source=self.name)
            inner_messages.append(msg)
            print("msg:",msg.content)  # Print for real-time countdown
            #await asyncio.sleep(1)
        final_msg = TextMessage(content="Done!", source=self.name)
        inner_messages.append(final_msg)
        print("final:",final_msg.content)
        return Response(chat_message=final_msg, inner_messages=inner_messages)

    async def on_reset(self, cancellation_token: CancellationToken) -> None:
        pass

[ekzhu]

I think you are on the right track.

One thing I suggest is to implement the on_message_stream method and have the on_message method calling it. So you can stream it.

Another thing is this:

if not messages:
            return Response(chat_message=TextMessage(content="No number received!", source=self.name))

It is possible for the agent to be called repeatedly without new message, so depending on your scenario, you may want to handle that case.

[rubans]

Thanks Eric. I assume there is no other non LLM way to do this all via Team Chat? I need to spend specific messages from agent to agent ?

…

________________________________ From: Eric Zhu ***@***.***> Sent: 11 February 2025 08:58 To: microsoft/autogen ***@***.***> Cc: Ruban Siva ***@***.***>; Author ***@***.***> Subject: Re: [microsoft/autogen] Make a triage agent to make a decision which agent to call without using LLM in AutoGen 0.4 (Discussion #5467) I think you are on the right track. One thing I suggest is to implement the on_message_stream method and have the on_message method calling it. So you can stream it. Another thing is this: if not messages: return Response(chat_message=TextMessage(content="No number received!", source=self.name)) It is possible for the agent to be called repeatedly without new message, so depending on your scenario, you may want to handle that case. — Reply to this email directly, view it on GitHub<#5467 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AAJCSW6GTPUAGIIQP7C3K5D2PG3UHAVCNFSM6AAAAABWZIJF26VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTEMJTGY3DQNA>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[ekzhu]

Actually, maybe you should consider SelectorGrouoChat with a custom selector_func that doesn't use models.

Alternatively for more custom orchestration patterns we recommend looking to the core API.

[rubans]

Hi @ekzhu
I did have a look at that but even custom selector_func. the model_client is a mandatory param I thought?
https://microsoft.github.io/autogen/stable/reference/python/autogen_agentchat.teams.html#autogen_agentchat.teams.SelectorGroupChat

[ekzhu]

Yes, because the model is used as a fallback in case the selector function returns None. If the selector always returns a string then the model won't be used

So you can pass None for model -- I think. Though I haven't done that.

[rubans]

I will try it out in that case.

[rubans]

Ok, it does work with a dummy OpenAI model instead. Thanks will check it out.