[AutoGen Studio] Group Chat Support(Extend DSL fromTwo Agents to Also Support Group Chat)

[victordibia]

What

AutoGen studio provides a json specification (DSL) that gets instantiated (via the workflow manager) into autogen agents.
The current version has support for workflows with two agents. This issue is meant to track tasks required to fully support GroupChat.

Task Items

• Implement initial GroupChat DSL specification
• Extend Workflow Manager to load GroupChat DSL specification
• Extend UI to support GroupChat DSL
  • Load and view group chat agents
  • Add GroupChat support to the "New Workflow" experience.
  • Support export of workflow configuration [Autogen Studio Feature Request]: The ability to export agents and workflows as python code. #1287

[sonichi]

Please be aware of #857 which extends group chat.
There is also a plan to extend it afterwards. @joshkyh please keep @victordibia posted.

[afourney]

This is a good opportunity to rethink default. I suggest:

allow_speaker_repeat = False

last_n_messages = "auto"

Make sure all agents have good descriptions, distinct from system prompts.

[joshkyh]

allow_speaker_repeat = False

Erh are you guys talking about GroupChat default? If so, allow_speaker_repeat has defaulted to True for sometime now. What if:

1. Users put one agent into GroupChat?
2. Users put two agents into GroupChat one of which is the Web Browsing agent?

But there might be strong motivations for setting it to False that I'm not aware of, though! :)

[afourney]

The most common scenario outside of AutoGenStudio where where this comes up is with the UserProxyAgent. GPT-4 GroupChat has an extremely bad habit of spamming the UserProxy, constantly asking for a response, in cases where it returns a blank message (the default auto reply when there is no code block). It's probably the most common situation from which loops occur.

Setting allow_repeat_speaker = False is a quick and dirty solution to this.

More generally, allow_repeat_speaker now allows a list. Excluding the UserProxy from this list (while, for example, including the assistant and web browser), is an even better solution. It allows those agents to have multiple consecutive turns... and this is especially helpful for web browsing.

Ideally, better orchestration would just solve the problem.

[joshkyh]

The most common scenario outside of AutoGenStudio where where this comes up is with the UserProxyAgent. GPT-4 GroupChat has an extremely bad habit of spamming the UserProxy, constantly asking for a response, in cases where it returns a blank message (the default auto reply when there is no code block). It's probably the most common situation from which loops occur.

Setting allow_repeat_speaker = False is a quick and dirty solution to this.

More generally, allow_repeat_speaker now allows a list. Excluding the UserProxy from this list (while, for example, including the assistant and web browser), is an even better solution. It allows those agents to have multiple consecutive turns... and this is especially helpful for web browsing.

Ideally, better orchestration would just solve the problem.

Thanks for the clarification!

[drosini-ibm]

Currently in Autogen Studio when I have a Session that is using a GroupChat Workflow, the output message is just the question that I originally asked. Would this roadmap item address that or am I having a different issue ?

[victordibia]

Hi @drosini-ibm ,

What version of autogenstudio do you have? The current version is v 0.0.56

To upgrade

pip install -U autogenstudio

[CrosswiredStudios]

I see many tutorials online showing a group chat workflow option, but the latest version seems to not have it available. Is this a known issue?

[victordibia]

@CrosswiredStudios , group chat worfklow is definitely still supported.
The user experience for groupchat has been updated (see video below)

• Build -> Agents -> new agent : create the agents you want to use in a group chat (user proxy or assistants)
• Build -> Agents -> new agent : create a group chat agent, click create. On the agents tab, add the agents you want. Hint: if your agents will be executing code, ensure you also add a user proxy or any agent that has code execution. Remember to add a model to the groupchat agent (needed for next speaker selection).
• Build -> workflow -> new workflow -> autonomous chat -> add a user proxy as initiator, and add your groupchat agent (that has other agents in it) as the receipient.

groupchat_1080.mov

[CrosswiredStudios]

Thanks! Not sure this is more intuitive UX wise, but i suppose its still easier than writing the scripts.

[CrosswiredStudios]

Its not working. It seems like it might be an issue in autogen itself. I made a blank script that just has the tutorial example code and it fails because a random agent "assistant" chimes in with a blank message.

https://microsoft.github.io/autogen/docs/notebooks/agentchat_groupchat

[2024-07-12 13:08:33.013] [INFO] [LM STUDIO SERVER] Processing queued request...
[2024-07-12 13:08:33.014] [INFO] Received POST request to /v1/chat/completions with body: {
"messages": [
{
"content": "You are in a role play game. The following roles are available:\n User_proxy: A user that can run Python code or input command line commands at a Linux terminal and report back the execution results.\nCoder: A helpful and general-purpose AI assistant that has strong language skills, Python skills, and Linux command line skills.\nProduct_manager: Creative in software product ideas..\n Read the following conversation.\n Then select the next role from ['User_proxy', 'Coder', 'Product_manager'] to play. Only return the role.",
"role": "system"
},
{
"content": "Find a latest paper about gpt-4 on arxiv and find its potential applications in software.",
"role": "user",
"name": "User_proxy"
},
{
"content": "Read the above conversation. Then select the next role from ['User_proxy', 'Coder', 'Product_manager'] to play. Only return the role.",
"role": "system"
},
{
"content": "",
"role": "assistant"
},
{
"content": "You didn't choose a speaker. As a reminder, to determine the speaker use these prioritised rules:\n 1. If the context refers to themselves as a speaker e.g. "As the..." , choose that speaker's name\n 2. If it refers to the "next" speaker name, choose that name\n 3. Otherwise, choose the first provided speaker's name in the context\n The names are case-sensitive and should not be abbreviated or changed.\n The only names that are accepted are ['User_proxy', 'Coder', 'Product_manager'].\n Respond with ONLY the name of the speaker and DO NOT provide a reason.",
"role": "system"
}
],
"model": "lmstudio-community/Meta-Llama-3-8B-Instruct-GGUF",
"stream": false
}
[2024-07-12 13:08:33.014] [ERROR] [Server Error] {"title":"'messages' array must only contain objects with a 'content' field that is not empty"}

[victordibia]

This is resolved.