FSM/Speaker Transition Control in Autogen v0.4

[aayushmittal2]

[Discussion]

Hi everyone,

I've been migrating from Autogen v0.2 to v0.4 and noticed that the FSM-like functionality for controlling agent transitions seems to have changed significantly.

In v0.2, we could control agent transitions using GroupChat with parameters like:

group_chat = GroupChat(
agents=agents,
messages=[],
max_round=25,
allowed_or_disallowed_speaker_transitions=graph_dict,
speaker_transitions_type="allowed",
allow_repeat_speaker=allow_repeat_speakers
)

This allowed us to define explicit state transitions between agents using the allowed_or_disallowed_speaker_transitions parameter.

However, in v0.4, I notice that RoundRobinGroupChat doesn't seem to have this functionality. Looking at the v0.4 documentation, I can't find any equivalent way to control agent transitions.

Questions:

1. Is there an alternative way to implement FSM-like agent transitions in v0.4?
2. If not, what's the recommended approach for controlling which agent can speak to which other agent in a group chat?
3. Are there plans to bring back this functionality in future releases?

My use case requires strict control over which agents can communicate with each other in a workflow, similar to a state machine. Any guidance on how to achieve this in v0.4 would be greatly appreciated!

Thanks!

[jackgerrits]

RoundRobinGroupChat selects speakers in a round robin fashion, as in cycling through each speaker.

What you're looking for is called SelectorGroupChat. In this the next speaker is selected via a function. By default this function is what you'd traditionally think of as group chat, where an LLM selects the next speaker. But you can pass a custom selector function (the selector_func arg) which can be used to implement the specific functionality you describe.

The second example in the docs for SelectorGroupChat show a custom function being used. The function is passed the conversation history and the next speakers name should be returned. You can determine who spoke last by looking at the source field of the most recent message.

[ekzhu]

Like what @jackgerrits has commented. See the selector group chat.

It's important to note that both v0.2 and v0.4 group chats are functionally the same: all participants share the same message history and messages are broadcasted to all participants. So there isn't a notion of "an agent speak to another agent", as an agent always speak to all. The graph in v0.2 specifies the order of speakers.

If you want strict sequences and do not want agents to broadcast, consider a sequential workflow using the Core API.