Why interaction, not diversity, drives creative output

There’s a comfortable assumption in multi-agent creative systems: if you want richer output, add more different agents. Diversity in, diversity out. I’ve come to think that’s only half right, and the less important half.

Diversity is a precondition, not a cause

A population of agents with different tastes, trained on different data, prompted with different personas — that’s the raw material. But if each agent generates in isolation and you pool the results, you mostly get a wider scatter of the same quality. More variety, not better work.

What changes the picture is interaction: letting agents perceive what the others made and respond to it. Agent A makes something; agent B hears it, reacts against it, pushes somewhere A wouldn’t have gone; A hears B’s reaction and adjusts. The output is no longer a sample from a fixed distribution — it’s a trajectory through one that the agents are reshaping as they go.

The measurable version

This is testable. Run the same population two ways: once with each agent generating solo, once with each agent perceiving the others between rounds. Hold everything else fixed. If diversity were the driver, the two conditions would look the same. They don’t. The interacting condition produces work that neither agent would have reached alone, and you can watch the influence flow — who leads, who absorbs, who resists.

Why I care

If interaction is the lever, then the design problem shifts. You stop asking “how do I get more different agents?” and start asking “what’s the shared substrate they interact through, and how rich is the channel between them?” That’s a much more interesting question, and it’s the one most systems skip.