How hard would it be to run a simulator with multiple LLMs. Say, one as the boss and a few as employees. Just let them talk, coordinate, and "work"? Could be the fastest way to test what actually happens when you try to automate management.
Multiple projects for autonomous multi agent teams already exist.
Left to their own devices, the LLMs would probably design a pocket watch.
This is quite literally what we've built @ Gobii, but it's prod ready and scalable.
The idea is you spin up a team of agents, they're always on, they can talk to one another, and you and your team can interact with them via email, sms, slack, discord, etc.
Disclaimer: founder