Multi agent type support

[fatihozturkh2o]

• More about AutoGen GroupChat: https://microsoft.github.io/autogen/docs/tutorial/conversation-patterns#group-chat

• Introduces new agent flow along with new agents (see the diagram above)

  • New agent flow can be the base for the future specialized agent orchestration
  • In the current flow, all the 'If' conditions and checks are piled up in CodeWriterAgent system prompt and weaker models easily struggle with this a lot of ifs.
  • With this flow, agents are aware of each others' presence and what they are capable.
  • Flexible to introduce new agents.
  • MainGroupChatManager decides when to select the next agent until the user request is met.
    • Or direct control over which agent to talk (in case it's needed, the manager can pick next agent via user instruction)
  • For now, as an alternative to Code Group, there is only Chat Agent that is only capable of simple chatting. If you ask questions like 'who are you', 'give me a joke' only that agent is expected to be picked up instead of going through code generation and execution steps.
  • For multi_agent type, I'm also bringing customized 'chat_history summarization'. As an example, if it's only chat_agent answer in the chat_history, then the last message will be the summary. If there were multiple agent interactions in the chat_history, then summarizing the process via LLM with a customized summarization prompt of mine.
  • Also added one more instruction in code_writer_agent system prompting to solve the following: In my tests, realized that sometimes, after code execution is completed, code writer agent wraps code output with output, and hence, code execution agent thinks there is a new code block and tries to execute it and fails. Later on, this runs into a loop.

[pseudotensor]

looking great!

[pseudotensor]

Want to merge so we can move on and add my changes we discussed. Let's make new issue for the cost issue.