GroupChatManager Not Respecting human_input_mode Settings Other Than ALWAYS

[anu43]

Hi all,

I'm experiencing an issue with the GroupChatManager where it doesn't seem to recognize the human_input_mode option as expected. Specifically, I can only trigger an interruption in the group chat by setting the UserProxyAgent to ALWAYS. When I set human_input_mode to other values like TERMINATE on the GroupChatManager, it doesn't behave as intended.

Below is a simplified version of my setup:

# Some other agents
# ...

# An admin
admin = UserProxyAgent(
    name="admin",
    human_input_mode="NEVER",
    llm_config=LLM_CONFIG,
    system_message="You are the Admin. You make the last call whether the given task is completed. If so, just say TERMINATE.",
    description="Makes the last call whether the task is done.",
)

# Construct the GroupChat
group_chat = GroupChat(
    agents=[admin, project_manager, tech_project_manager],
    messages=[],
    max_round=10,
    speaker_selection_method="auto",
    send_introductions=True,
    select_speaker_auto_llm_config=LLM_CONFIG,
)

# Create the GroupChatManager
group_chat_manager = GroupChatManager(
    groupchat=group_chat,
    human_input_mode="TERMINATE",  # Expecting to interrupt on TERMINATE commands
    system_message="You are a group chat manager. Your job is to route between the agents depending on their expertise and responsibilities.",
    is_termination_msg=lambda msg: "TERMINATE" in msg["content"],
)

During execution, the following log indicates that the chat manager routes messages and processes a TERMINATE command:

--------------------------------------------------------------------------------

Next speaker: admin

admin (to chat_manager):

The task of providing a blog post on how the transformer architecture is evaluated until today has been completed with a detailed and technically improved prompt. TERMINATE.

--------------------------------------------------------------------------------
**********
ChatResult(chat_id=None, ...

My Question:
Am I missing a configuration point or misunderstanding how human_input_mode is intended to work with the GroupChatManager? Any insights or suggestions for resolving this issue would be greatly appreciated.

Note: working in v0.2. I will try the v0.4 at some time.

Thanks in advance for your help!

[ekzhu]

The GroupChatManager doesn't work with human input mode, you need set this to a user proxy agent and have it initiate a chat with the group chat manager.

[anu43]

In the UserProxyAgent, when setting human_input_mode={ALWAYS or TERMINATE}, should the initialization remain as shown below?

# Start the chat
chat_result: ChatResult = admin.initiate_chat(
    recipient=group_chat_manager,
    message=PROMPT,
    summary_method="last_msg",
)

Would this configuration still align with the intended behavior, or are there specific adjustments required in this context?

[abhinavpgagi]

@staticmethod def get_user_proxy_agent(agent_details: Dict[str, Any]) -> UserProxyAgent: return UserProxyAgent( name=agent_details.get("Name", "default_user_proxy"), description=agent_details.get("Description", "A human user proxy"), input_func=None )
How do i set human_input_mode to "NEVER" in the new version of autogen v0.4

[ekzhu]

If you don't want to have human input, you just don't use the user proxy agent.