Replace task executor with a SmartGPT scheme
xtalax opened this issue · comments
So now this has been demonstrated:
This looks like a good scheme to use for the task executor
Hi @xtalax
This is relatively simple to implement using the existing agents we have. The 3 Output variables would be represented by separate calls to the execution agent. The Reflexion step can also be performed by the execution agent using a different persona. The final step would be performed by the status agent. It's just a slightly different prompt.
Our system is intended to be a framework where you can quickly design cognitive architectures like this rapidly. Once we implement superclasses for agents, I think adding the ability to select the persona being used as a parameter might simplify this process and make reusing agents more intuitive. What other features do you think would help you rebuild this logic more easily?
We are going to need some kind of markup language (Maybe yaml?) to define different architectures and modified prompts, but then does this just become langchain with a default agent library?
Documentation of the agent model and interface would be the biggest leap forward at the moment IMO
EDIT: This idea of multiple generation and refinment can also apply more generally to other processes to improve the "intelligence" of the action