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

This paper asks a well-posed and practically important question: do persona prompts meaningfully separate the coordination behavior of multiple copies of the same LLM, or does the diversity collapse when agents know they are interacting with each other?

The answer is clear. Persona diversity is real when each agent picks freely: between-persona agreement on abstract menus starts at 0.776. Under a coordination frame, it rises to 0.916, removing 62% of persona disagreement. The gain comes almost entirely from agreement, not from agents becoming individually more decisive. Personas converge near where the default model sits, not toward a human-style Schelling focal point. This is a meaningful result for AI oversight: using multiple persona-prompted copies of the same model is not a reliable defense against zero-channel tacit coordination.

On pricing, the finding is more nuanced. Brand identity survives every frame: the premium assistant always prices higher than the discount one. But the tier gap narrows by 30% under coordination, and the brand ordering is preserved only imperfectly. The paper frames this correctly as partial compression rather than full convergence.

The mechanistic analysis is one of the paper's strengths. A shared framing direction in the residual stream at layer 32 is decodable with AUROC 0.9995 and fires for frames that produce little genuine convergence too (the hidden-rule frame). Reduced attention to persona tokens under coordination causally accounts for about a seventh to a fifth of the agreement gain. The attention knockout is carefully controlled against equal-length template spans and user text. The paper does not oversell this: the remaining four-fifths of the mechanism is not yet located.

The main limitations are the single-model constraint (Qwen3-32B with reasoning off), the single-shot forced-choice design, and the abstract-menu setting for most results. The pricing task has nine agents and thinner statistical power. The paper is transparent about tests that fail or are underpowered, which makes the positive results more credible.

This is a thoughtful and technically strong project that tests an important assumption: whether giving the same model different personas actually reduces its ability to coordinate. The methodology is a major strength, particularly the sampling-free measurement and the separation of confidence from agreement, which could be useful beyond this study. The controls, pre-registration, multiple-testing correction, and transparent reporting of failed predictions also give confidence in the research process. The main limitation is that the headline agreement effect is relatively small in practical terms, while the more deployment-relevant greedy-decoding result deserves greater emphasis. Coordination is also explicitly instructed, reasoning is disabled, and most experiments use abstract tasks where personas have limited practical significance. Some of the stronger robustness measures were developed after the initial results, and the attention explanation is partly confounded by prompt length. A follow-up using reasoning-enabled models, multiple model families, uninstructed coordination, and fully length-controlled experiments would make the findings much more relevant to real-world multi-agent deployments.

A built instrument used to answer a question that practitioners actually ask: does giving copies of one model different personas keep them from coordinating? With Qwen3-32B, 23 personas and a million scored prompts the answer is that personas differ when each agent picks what it likes.. When any frame involves another agent’s choice 62% of between-persona disagreement on abstract menus and 44% on pricing is removed. This happens through agreement, not confidence. It works about equally for human and AI partners. Telling an agent that its partner is a copy makes personas slightly more alike.

The sampling-free readout with a decisiveness/agreement decomposition is reusable. The confirmatory tests are pre-registered with Holm correction and failures are reported in full. Remark 1. A dependence test at fixed context cannot see shared- coordination capacity. Is a point that the detection literature should absorb. The mechanistic section is appropriately modest.

Limits: the study uses one model, one shot, no reasoning and instructed coordination. The baseline persona diversity is already close to that of a chooser (0.776, vs 0.76) so the percentages removed are percentages of a small quantity. The abstract should state the baseline. The pricing task has no payoffs.

The abstract contains a sentence and typos. The required dual-use appendix is missing. The reasoning-on version is the experiment that matters for deployed agents.