I get super excited about AI agents and how they can work together to solve problems so here’s something cool I’d like to share: AI will soon be able to work together in huge numbers—way more than humans can. And these are more than just bots that get in between you and a human customer service agent.
There's new research showing that Large Language Models, especially the bigger and more advanced ones, can collaborate in groups of 1,000 or more. That’s interesting because according to Dunbar's theory, the maximum number that people can have meaningful relationships with at any one time is about 150. This includes coworkers like the kind you’d have collaborating on a project with you.
Dunbars number means that there’s a limit for how many people you can have a meaningful relationship with. So even a well-structured team that gets too big will start to fall apart communication and it starts to get harder to get anything done.
But for AI, maybe there is no limit on how it could collaborate.
The Experiment
In this study, they decided to test if AI models—specifically large language models (LLMs)—had limits like we do when it comes to group size. So, they created multiple instances of the same model, each one given a different random opinion. The twist was that these AI “agents” could look at the opinions of their peers and decide if they wanted to update their own opinions or not.
The goal was to see how well these AIs could come to a consensus—essentially, whether they could agree on something. Smaller or older models like Claude 3 Haiku or GPT-3.5 Turbo couldn’t reach a consensus when put in larger groups. They sort of broke down.
But the big guns like the 70-billion parameter version of Llama 3 and GPT-4 Turbo were a different story. GPT-4 Turbo got 1,000 agents to work together and agree on something—without any signs of breaking down or hitting a limit.
So how is that even possible? Well, AIs don’t have the same social or biological limitations we do. They don’t get tired, lose track of conversations, or get bogged down by emotions and distractions. They process information super fast and aren’t constrained by the same bandwidth issues we have when trying to juggle multiple interactions.
Why Does This Matter?
Now you might be thinking, “Okay, cool, but why does this matter?” Fair question. What’s interesting here is if AI can collaborate on a massive scale—bigger than human teams ever could—it opens up some pretty big possibilities.
AIs working together on massive complex problems—coordinating global logistics, climate modeling or even managing super intricate financial markets. When humans do this kind of stuff, things get messy because of our limitations—miscommunication, cognitive overload and decision fatigue. But if AI agents can scale up and work together in these big groups, it could make some tasks much more efficient.
AI Consensus Power--And Limits
Okay, but before we get too hyped, there’s a catch. Just because a group of AI agents can reach a consensus doesn’t mean they’re coming up with the best answer. I mean think about it—if you clone a hundred copies of yourself and all of you have the same biases, you might all agree on something but that doesn’t mean it’s the right call! Philip Feldman from the University of Maryland said as much, “It’s cool to see the models agree quickly but the solution they land on might not be optimal.”
And there’s the cost of computation. Running a thousand instances of an AI model isn’t cheap. So even though these experiments show AI can theoretically collaborate on a massive scale, it might not always be practical with today’s resources.
OpenAI's Swarm Framework
OpenAI is working on something called Swarm, which is a way to build and deploy multi-agent systems. If you’ve heard of these “multi-agent systems” before, they’re setups, where multiple AI models (or agents) are given tasks. They can work together, transfer responsibilities or handle different parts of a bigger goal.
Swarm is cool because it’s a lightweight way to manage these agents so developers can orchestrate how the AI models interact. It’s still experimental but OpenAI is open sourcing it so developers can play with it. The framework is designed to be flexible and controllable with a couple of core concepts: Agents (holding instructions and tools) and handoffs (one agent passing control to another).
The Future of AI Collaboration
With the research showing AI can collaborate on a massive scale and tools like Swarm making it easier to manage multiple AI agents, it feels like we’re on the verge of some big changes in how AI is used. If AI can work together in groups of thousands—far more than humans can manage—we could see everything from large-scale problem-solving to massive automation projects that go way beyond what’s possible today.
But as always, remember collaboration doesn’t always mean best results. A group of AI agents might agree fast but whether it is the right solution is a whole different story. That said, we’re heading into a future where AI can scale up and collaborate and open doors we haven’t even thought of yet.
What do you think—would you let 1,000 AI agents work for or alongside you?