Final Response Agent - Githubissues

This PR is more like an idea placeholder for a Final Response Agent to generate guided final responses after each agent conversation terminates. (for 2agent type)
Requires more testing with OSS as well for sure
Main motivation working on this agent is, quite frequently, I'm still seeing the issue where final agent response includes words like 'Thank you for running the code!' or 'The code execution was successful' etc. Or, when the user asks just for a code snippet, the final answer doesnt include the code at all, which is what user is actually interested in.
I've started playing with changing <no_code_executed_notes> and actual code_writer_agent system_message to instruct the llm to avoid such mentions, however even Claude 3.5 doesnt follow these instructions for some reason. (Need to play with this idea more) So, based on my first exploration, I find it easy to instruct llms via this extra final response call with a small system_message. We might consider having this final step and more control over the final responses at the cost of having one extra llm call 🤔
Again, this is more like a placeholder rn, needs more testing + we need to make sure that the method I mentioned one-point above doesnt work at all.

Note that the final code_writer_agent still says Thank you or 'execution was successful'. Or doesnt put code snippets in the final message at all https://github.com/user-attachments/assets/adffe394-b969-4aa2-8ac7-7fb1d83fee1d
In this PR, I've introduced new Final Response Agent which takes the previous agent conversations as llm messages, crafts a final response. It should be paying attention on:
- The final response to sound as if talking to the main user directly for the first time
- Writing down the new message based on the information avail in the chat history only
- Answering main user request first
- Sharing interesting and relevant findings from the internal agent conversations

h2oai / h2ogpt