When Agents Consider Each Other, Personas Fade: Coordination Capacity Among Persona-Prompted LLMs

Imagine an online marketplace with 3 sellers where each run an AI agent to set prices. One is a premium brand, one a discount store and one an eco-friendly label. But all three agents are copies of the same AI model behind different instructions. They never exchange a single message, yet they can still end up pricing similarly. Copies of a model tend to make same choices in same situations, which means they can coordinate even without any channel that an auditor could inspect or cut off. The usual reassurance is that giving each agent its own persona, with different goals, values and brand identity, will keep their decisions apart.

We tested that reassurance on Qwen3-32B with 23 personas. We used simple choice games and a pricing task, and read off each agent’s full probability of picking every option, so how likely two agents are to match can be computed exactly. When agents simply choose what they like, the personas do differ. Once the decision involves another agent, whether matching a partner, avoiding one or guessing what someone else picked, the differences shrink sharply. A coordination prompt removed 62% of the disagreement between personas, and the chance that two of them gave the same answer rose from 40% to 50%. The agents converged on a shared answer close to what the model gives with no persona at all. Describing the partner as a person rather than an AI made little difference, and telling an agent its partner was an exact copy of itself did not bring its persona back either. On pricing, the premium and discount agents kept their order, but the gap between their prices narrowed by 30%. Inside the model, the answer pays about a fifth less attention to the persona instructions once another agent is involved.