Exploring How Multi-Agent Systems Develop Their Own Language

When we think about artificial intelligence, we often imagine smart machines that can talk to humans. But what happens when AI systems need to talk to each other? In my recent research, “To What Extent Do Multi-Agent Systems Use Compositionality in Multi-Agent Interactions?”, I explored how AI agents develop and use communication to cooperate on shared goals.

I trained AI agents using a speaker-listener setup where one agent had to describe a goal, and the other had to act on that information. These agents communicated through short, discrete messages, learning entirely from trial and error. The goal was to see if their invented communication showed signs of compositionality—the ability to build complex meanings from simpler parts, much like human language.

The results were fascinating. The agents developed structured communication patterns, with measurable topographic similarity (ρ = 0.69), showing a clear relationship between messages and their meanings. Their communication significantly improved teamwork, raising performance scores from -150 to -87. However, the mutual information between the agents’ actions (I(S;L) = 0.017 bits) suggested that while their “language” was organized, it was not yet fully compositional. In other words, the agents had learned to communicate consistently, but not creatively.

Disabling communication caused performance to drop sharply, confirming that their messages were essential to success. This means that even simple AI systems can invent useful communication protocols that support teamwork. Yet, the messages they use still lack the flexibility and productivity we see in human language.

This study offers a step toward understanding how communication might evolve in artificial systems. Future work could explore richer environments, larger vocabularies, and longer training to encourage deeper linguistic structure. The findings help show that while today’s multi-agent systems can “talk,” they are still far from mastering the art of language.