-
How to prevent the agent from leaving the loop in between tool calls, saying it’s “going to move on to the next step”, or to ask the user for confirmation.
-
How to make the agent manage it’s tokens autonomously and intelligently, by identifying and discarding messages or context that are no longer relevant.
(Especially important for when there’s this one gigantic piece of content among the messages that keeps getting sent back in each new API call, making the token count grow exponentially)
- One path that I’m exploring is to have the one task or step in a different agent, so the main agent calls the agent that executes this task, so the context stays separate.
-
On that last note, to understand the balance between having 1 agent with a huge prompt and a lot of tasks/steps, or one “manager” agent that calls multiple small agents with small prompts and single tasks.
-
When asking the model to complete a task, what’s the perfect balance between asking it do it a certain way (step by step) as opposed to just letting it decide by itself?
Broad vs. specific prmpts - are there cases when one performs better than the other?
As of now, I feel the latter is the best choice in general, but I wonder if that’s gonna change as the models get better. Or if that’s even true for all topics and types of tasks.