I enjoy reading through the verbose logs of the agents communicating with each other. One thing that would really provide a lot of value would be to have some sort of visualization that shows the communication between the agents.
Does something like this exist? It would make it highly valuable when demoing crewAI apps to clients.
@coderberry The closest you can get is currently plotting a flow. Although thatās probably not what you want. You can plot only the workflow, but not conversations.
there was this older agent software called chatdev that used little pixelated characters that were in a little office and you would see them interacting with each other to accomplish the task, it was pretty cute
Iāll soon be proposing a fully automated, agent-based system to a client, and Iām finding that the concepts and capabilities are challenging to grasp. These types of systems feel so futuristic that, at least for me, using visuals and relating the flows to human interactions makes the ideas easier to understand.
This is one of the reasons I was drawn to CrewAI. Itās easier to understand what building a crew is vs an agentic AI system.
Yes, sort-of. I have the crew set up to return not only the response message but their thought process in their responses and then use the step callback to send the output to a slack thread. It seems to work pretty good.
@coderberry
In a similar vein
A āhackyā way of accomplishing this might be to have a Task/Agent dedicated to capturing the conversations. You could then your have final Task consolidate everything into the final output. From there, youād have easier access to the conversations. Obviously this is a possible extra API to an LLM (if not local), so a more cost-efficient method would probably be in some custom non-AI logic.
I considered that as well. The value for me is to understand the thought processes for each agent in the process. Using slack for visualizing this is pretty nice with their block kit framework