[Bug]: max_consecutive_auto_reply not working on recipient agent

[marklysze]

Describe the bug

The max_consecutive_auto_reply parameter on an agent is not terminating the chat when it's on the agent that is the recipient in initiate_chat.

In this example it should terminate when my_agent goes to reply for a second time, but doesn't:

from autogen import ConversableAgent
llm_config = {"model": "gpt-4o-mini"}

# Doesn't terminate on auto reply
my_agent = ConversableAgent(
    name="helpful_agent",
    llm_config=llm_config,
    system_message="You are a poetic AI assistant, respond in rhyme.",
    human_input_mode="NEVER",
    max_consecutive_auto_reply=1
)

other_agent = ConversableAgent(
    name="other_agent",
    llm_config=llm_config,
    system_message="You are a scientific AI, respond with science facts.",
    human_input_mode="NEVER",
)

other_agent.initiate_chat(my_agent, message="What's the big deal about AI?", max_turns=10)

Whereas, if we change the agents around in initiate_chat, it does:

from autogen import ConversableAgent
llm_config = {"model": "gpt-4o-mini"}

# Does terminate on auto reply
my_agent = ConversableAgent(
    name="helpful_agent",
    llm_config=llm_config,
    system_message="You are a poetic AI assistant, respond in rhyme.",
    human_input_mode="NEVER",
    max_consecutive_auto_reply=1
)

other_agent = ConversableAgent(
    name="other_agent",
    llm_config=llm_config,
    system_message="You are a scientific AI, respond with science facts.",
    human_input_mode="NEVER",
)

my_agent.initiate_chat(other_agent, message="What's the big deal about AI?", max_turns=10)

Steps to reproduce

See code above.

Model Used

Not model specific, used gpt-4o-mini

Expected Behavior

Terminated before replying a second time.

Screenshots and logs

No response

Additional Information

No response

[marklysze]

This is likely a broader issue, I encountered the same bug if I register a reply function with an agent that should terminate upon reply, but it doesn't if the agent with the reply function is the recipient in initiate_chat.

This should terminate when it's agent_b's turn to reply, but it doesn't.

from typing import Any, Dict, List, Optional, Tuple, Union
from autogen import ConversableAgent, Agent
llm_config = {"model": "gpt-4o-mini"}

a_agent = ConversableAgent(
    name="agent_a",
    llm_config=llm_config,
    system_message="You are a poetic AI assistant, respond in rhyme.",
    human_input_mode="TERMINATE",
)

b_agent = ConversableAgent(
    name="agent_b",
    llm_config=llm_config,
    system_message="You are a scientific AI, respond with science facts.",
    human_input_mode="NEVER",
)

def my_reply_func(
    recipient: ConversableAgent,
    messages: Optional[List[Dict]] = None,
    sender: Optional[Agent] = None,
    config: Optional[Any] = None,
) -> Tuple[bool, Union[str, Dict, None]]:
    return True, None

b_agent.register_reply(
    trigger="agent_a",
    reply_func=my_reply_func,
    position=0

)

# b_agent being recipient doesn't work here, switch a_agent and b_agent around here and it works.
a_agent.initiate_chat(recipient=b_agent, message="What's the big deal about AI?", max_turns=10)