When you are working with LLMs it can be helpful to think of them as people instead of as computer programs. This framing can be helpful because they are more like an employee than an Excel sheet, even though the truth is that they are their own kind of thing.
One thing LLMs don’t have much of as of 2023 is agency; they passively await for you to issue a prompt. That is changing, we’re approaching a world where we might say, “Monitor the business news and let me know if anything relevant is happening. Send our queries if you need more information and if you need to do any financial analysis hook up to one of the relevant tools for that.”
The other thing about LLMs in 2023 is that we tend to deal with just one at a time. That is changing. We are approaching a world where we will build swarms of AI Agents, and each one might focus on a different thing, e.g. one for marketing, one for finance, one for logistics and so on. In fact, we won’t necessarily be doing all the creating, we will start with an agent, let’s call it the AI-CEO and tell them to create sub-agents as needed.
This leads us to a world where we have a hierarchical swarm of agents which is very much like a human organization minus the humans. David Shapiro, who is leading much of this work, said something interesting. He said building the agents is relatively easy, what’s hard is deciding who talks to whom, how much and how often (see this video).
I might add to that “How do you set goals?”, “How do you allocate resources between agents?”, “How do you resolve disputes?” “What do you do if you suspect an agent isn’t up to the job and you need a new one?”
In other words, the whole messy world of management will reappear in this world of AI agents.
What is exciting here is that management thinkers may be able to help guide the AI teams on how to organize AI agents into effective teams and organizations. It’s also possible that experimenting with groups of AI agents will provide insights into how better to manage human organizations. It’s an opportunity for in silico experimentation, something I mused on briefly at the end of my book (Management for
Scientists and Engineers) with the hope that sometime in the future we might have a platform that could do this in silico work. Seems like, much to my surprise, that platform is already here.
LLMs are sufficiently different from humans that we may find that useful parallels between organizations full of humans and those full of AI agents are few and far between. Whatever the case, there is no doubt we’re heading into a very interesting world.
To learn more check out Shapiro videos, such as this one: