AutoGen Studio: why can we only run a "single team"?

[joslat]

What feature would you like to be added?

IMHO the name is wrong.
I am running an agentic workflow which can be composed of a workflow which can contain other workflows (subchat) or in the future a mix of business workflow and agentic workflow which can contain sub-workflows (n levels of nested chats, for example)

So calling it "team" imho is wrong. We should call it workflow which enables composability. when we talk about a team, we do this in singular so it implies no composability.

Why is this needed?

make autogen studio more clear and future proof.

[victordibia]

Hi @joslat ,

I agree with idea of workflow being relevant to many business and usage scenarios as you mentioned and the benefits of implied composability.

We have mostly used the term "team" as a reference to "multiple agents" collaborating autonomously and with humans to solve a task . Teams can be composable too - composed of multiple agents, and agents themselves could wrap teams (see the society of mind agent present)

Note that this is only in the high level AgentChat API where we offer some presets (a team being one of those presets). In the low level core library, developers can implement their own workflows

I agree that we should also support more structured workflows (see #4623 ) and work in that terminology and capabilities more carefully into the library!

@joslat , do you think enabling the ideas in #4623 will address the workflow angle ?

Great feedback!

@ekzhu @afourney @gagb

[ekzhu]

Thanks @joslat for pointing this out. We are exploring the direction of workflows in studio and in that case, it will allow user to compose multiple teams. Currently, you can create sub teams using custom agents or society of mind agent as Victor described.

[joslat]

@victordibia yes, #4623 is a must imho.
If we think in metaphors, we usually see a team of up to certain number of components, as things are evolving, I imagine we can reach teams of 10, 20, 50 and 100 agents collaborating soon in a "team" - I mean this year. While technically is correct by forllowing the #4623 we are designing a graph of Agents and Tasks to achieve a goal which involves agents which create, collaborate, review and correct each other, and ensure quality.

I see this as a workflow designer, which can contain teams of agents (group chat, two agent chat, sequential chat, etc...).

And yes, I have in mind exactly something like the Society of Mind. There I would like to be able to do that as well as it is one of my favorites, even more when you invoke all the agents in parallel: https://github.com/joslat/SemanticKernelAgenticChatWorkflowsExtension/blob/main/AutoGen/AutoGenChatWorkflow_AskForFeedback.cs (based on an example from @LittleLittleCloud ) - this is super efficient, way more than a group chat with an agent chat manager supervising and doing "middleman" on every single interaction...

[joslat]

Thanks @joslat for pointing this out. We are exploring the direction of workflows in studio and in that case, it will allow user to compose multiple teams. Currently, you can create sub teams using custom agents or society of mind agent as Victor described.

Is there an example of this? I'd like to check this out then :)