There have been a few previous topics on how AI (large language models) could possibly be useful in implementing or teaching AI ( Sociocracy for All AI Bot , Can AI change sociocracy? ), but what about about the other way around?
I have been using coding “agents” quite heavily in the past 5-6 months, as they have now reached a place where they can be quite effect at the level of work I am doing in my PhD. Most other members of my lab have also started using them to help them code.
As I explore how much they can take on, I have begun using “sub agents” or having one agent launch other agents to help it. This kind of technique is helpful for dividing up work and also managing “context”, basically how much each agent can hold at once. If work is divided, then their memories can be more focused and they don’t have to swap things out as much.
Anyway, I started to think that maybe Sociocracy could be useful as a driving framework to help coordinate. I was curious if anyone else has started exploring this as well. Basically, the idea being to have the main agent the “general circle” and to move all critical work out of there into different “implementation circles” and then also have different expertise circles for pulling in different perspectives. Then have a mission circle that’s responsible for holding overall direction. This is useful when you have tasks that might run for hours at a time without human intervention, and you want to help guide the agent to stay in the happy path, and not get lost. In this case each “circle” is not a group, it’s just one agent, so maybe it would be more appropriate to call them roles, but I have found the division of labor metaphor to map OK.
I am just trying this out on some harder problems, so I can’t say for sure in any evidence based manner whether it’s useful, but it does seem like qualitatively it’s starting to make a difference. Of course agents != humans, but they are trained on humans, so it could make some sense that techniques we have developed for management also translate somewhat.
It could make an interesting research study to evaluate different types of management styles in these settings to see which ones are the most effective at solving longer open ended tasks!
Here is my (actively evolving) “skill” document that tells the LLM to work this way. Let me know if you try it out and have any feedback, or have any thoughts on the approach at all. I hope this isn’t heresy to try to apply these models for AI agents and to use all the knowledge, especially from the books, to help create these skills!