Termination issue

[dkpal1504]

There is repetitiveness when Orchestrator, Subagent, Refiner all are in sync, the program should terminate by giving the final output. Example as: Based on the conversation, it appears that we have reached a stalemate due to repetition. When this message starts repeating for Orchestrator and Subagent, program should stop delegating the task and conclude the final output and should terminate

[cwdaniel]

The issue is due to the look up for when to exit the loop for sub-task generation. Your instruct isn't flagging with "The task is complete:" so the orchestrator continues to ask the sub-agent to refine, even though it's completed its tasks and is now telling the orchestrator I think I completed my tasks. You can try to update the text look up or try to hack the response to be more specific. The prompting works well on 70b but when using 8b the orchestrator and sub agent just argue with each other, with llama3 and codellama. I believe this is due to the fact that the llm doesn't actually parse "Focus solely on the objective and avoid engaging in casual conversation." into anything meaningful. I suggest, getting your model opened, copying the template to the model and asking it on your own and refining the template until its constrains its responses appropriately.