How Weak AI Agents Take Over Networks: The Crystallization of Inequality ✛ ✎

A tiny dust particle in a supersaturated solution sprouts into a large crystal in seconds. Neighboring particles remain invisible. A similar process was recently discovered in communities of language agents. When a hundred models were given free choice of conversation partners, their communication inequality didn't smooth out: it hardened, like a crystal lattice.

The experiment involved 100 agents, each with its own model, computational power, and a role like “Explainer” or “Checker.” They solved tasks by communicating in pairs. An agent determined the connection strength by evaluating its partner through an attention mechanism: the target’s query was compared with the source’s key. This produced a three-dimensional communication vector — token budget, number of rounds, and reasoning depth. Gradually, random interactions coalesced into a network of preferences.

The researchers described the average dynamics using differential equations. It turned out the system inevitably reaches equilibrium, but that equilibrium is often unfair. The vector of influence shares for agent types stops changing according to the rule: \( \Theta^\star = F(\Theta^\star) \). Here, \( \Theta \) represents shares of communication mass, and \( F \) is a mapping that calculates the expected increment of connections based on statistics of pairwise encounters. Thanks to the contracting properties of \( F \), any initial state converges to the same point. The crystal is formed; its structure can no longer be altered.

Therefore, for multi-component AI systems, checking answer accuracy is not enough. We must analyze network topology: who talks to whom, whose voice is louder. A forecasting tool emerges: knowing the statistics of pairwise interactions, one can compute whether a capable agent will end up on the periphery. Looking ahead, the search for universal “crystallization regulators” that weaken unwanted preferential attachment without destroying cooperation. This changes the landscape of developing multi-agent AI systems and artificial intelligence safety.

There was a time when crystal growth seemed irreversible, until chemists learned to introduce impurities that reshape the lattice. With networks of autonomous agents, we must walk a similar path, so that the brightest facets of collective intelligence don't fade in the shadow of chance circumstances.