new multi agent flow with terminate agent

[fatihozturkh2o]

• Better for prompts where user is only asking for the code but not for the execution.
  • Now GroupManager decides whether code execution is needed or not based on # execution: VALUE tag coming from code generator agent.
• Now Group manager has capability of terminating the chat as soon as user request is met. It's done by selecting the next agent as 'terminate_agent'.
  • This termination process can be extended to detecting repetitive errors in the agents' conversation. If Group manager decides that there has been repetitive/unsolvable errors in the chat history, it can decide to terminate the chat.

Based on current h2ogpte agents pytest, it shows similar results to the other agent types. The most distinctive success comes with the question "Write a code to show an image in matplotlib. Do not execute any code." Because this new flow is the only flow that does not even call 'code_execution_agent'.

• testing

[RahulVadisetty91]

Ensure that executor is properly defined and compatible with the expected configuration. Verify that the configuration for code_execution_config aligns with the new or updated requirements for code execution.

[pseudotensor]

Write a code to show an image in matplotlib. Do not execute any code. works normally for the 2 agent case. It does not execute the code.

[fatihozturkh2o]

Write a code to show an image in matplotlib. Do not execute any code. works normally for the 2 agent case. It does not execute the code.

Yeah, but, with 2agent, the code execution agent is called anyways like in this example: (even code generation is called once more though it's not needed)

2agent

Screen.Recording.2024-09-17.at.12.48.56.AM.mov

I'm aiming not even calling code_executor agent with the new flow (because # execution: false information is out there anyway, thought the manager could make use of it during agent selection)

PR multiagent

Screen.Recording.2024-09-17.at.12.54.53.AM.mov