TextMessageTermination Condition

[ekzhu]

Confirmation

• I confirm that I am a maintainer and so can use this template. If I am not, I understand this issue will be closed and I will be asked to use a different template.

Issue body

A termination condition that stops the team when a TextMessage is emitted from a specific source, or any source.

For example, to terminate when a specific agent produced a TextMessage:

TextMessageTermination(source="assistant")

To terminate when any agent produced a TextMessage

TextMessageTermination()

• In Swarm group chat team, it is useful to terminate the team when an agent no longer hands off or producing tool call results. AgentChat Swarm to support forced handoff to avoid agent getting stuck #5611 (comment)
• In a single agent scenario, often useful to repeatedly call the same agent in a RoundRobinGroupChat until it stops producing tool call results. Gemini models via Openrouter not supported #5621 (comment)

For example

assistant = AssistantAgent(...)
team = RoundRobinGroupChat([assistant], termination_condition=TextMessageTermination())
result = team.run(...)

As part of this issue, also update the documentation:

• https://microsoft.github.io/autogen/dev/user-guide/agentchat-user-guide/tutorial/termination.html
• Add a new subsection "Single Agent Team" before this subsection: https://microsoft.github.io/autogen/dev/user-guide/agentchat-user-guide/tutorial/teams.html#creating-a-team. And show the example above.
• Add pointer from the agents doc to this example regarding the common question of how to run an agent repeatedly.

[EItanya]

I 1000% agree with this. In fact I think it could be awesome with SocietyOfMindAgent. I can create a quick PR?

[ekzhu]

@EItanya Thanks! Assigned to you.

[EItanya]

@ekzhu #5742

[EItanya]

I actually had a quick question about the SocietyOfMindAgent generally, I can also create another issue.

I noticed that the SocietyOfMindAgent does not support streaming chunks:

autogen/python/packages/autogen-agentchat/src/autogen_agentchat/agents/_society_of_mind_agent.py

Line 154 in dd0781a

if isinstance(inner_msg, ModelClientStreamingChunkEvent):

.

Our use-case "requires" streaming responses back from the sub-team, and probably doesn't require a final message output from the society of mind agent. Should we create a custom version without that, or potentially make the society of mind more configurable for that use-case?

[ekzhu]

The ModelClientStreamingChunkEvent is yielded but not added to the inner messages. So it should be in the stream.

SocietyOfMindAgent is a very opinionated implementation of nested chat. So I recommend creating custom agent for any serious nested chat scenario.

[EItanya]

Sounds good, I will do that. I still pushed a few fixes the the SocietyOfMindAgent in my PR to fix the declarative approach.