Best Way for Agent in Multi-agent Group Chat to Reference Output from Other Agents

[corytomlinson]

Consider Agent_A and Agent_B are part of a group chat like...

chat = AgentGroupChat(agents=[Agent_A, Agent_B])

...and let's say that Agent_A produces some output like...

{"key": "value"}

...and now when Agent_B runs I want Agent_B to reference "value" from the output of Agent_A from the chat.history, how is that best achieved?

It seems to work intermittently by just mentioning in the instructions for Agent_B to "read the group chat", but I wonder if there are more explicit ways to instruct agents how to access parts of the group chat history.

[sophialagerkranspandey]

Tagging @crickman

[crickman]

Each agent in the chat has explicit knowledge of the entire chat. I'd be interested to understand more specifically on your use-case.

[AyushAP8]

I have a termination and selection strategy built from a prompt, does the kernel assigned to these strategies have access to the chat history?

[corytomlinson]

It's clarity on this statement that I am looking for...

"Each agent in the chat has explicit knowledge of the entire chat."

What does that mean exactly, and in technical terms.

The use case is I want to provide explicit instructions to an agent to look for something another agent in the group chat provides in one of their messages. How do I do that? Do they see the xml? Do they see the parsed text? Do they see some other format?

As I mentioned I can instruct the agent to "read the group chat" and it works somewhat reliably. I want to be able to instruct the agent exactly where to look.

[corytomlinson]

@crickman any clarity you can provide here? Still struggling with this.

[smokhasi]

In a sequential execution of agents such as

AgentGroupChat(
        agents=[database_agent, research_agent...])

Is there a way to extract a portion of the chat history and pass it on to the next agent?

[TaoChenOSU]

Each agent in a group chat orchestration maintains its own chat history that includes the conversation of the group. When an agent is selected to talk, it will use its own chat history as context to generate a response. The response will be broadcasted to the other agents, who will then add that response to their chat history. Note that if an agent decides to use tools, the tool messages will not be broadcasted, because those messages are considered internal to the agent itself.

In summary, each agent maintains a chat history that has responses from other agents in the group + tool messages it has generated.