When the Team Becomes an Echo Chamber: Authority, Scale, and Topology

Multi-agent AI brainstorming changes the diversity problem because agents no longer generate in isolation: each contribution can alter the context available to the others. This chapter examines how model alignment, role authority, message visibility, group size, and communication topology shape that interaction, so the reader can tell whether a team of agents is adding independent information or merely producing a more elaborate consensus. The central finding is not that collaboration is inherently harmful. It is that collaboration can couple otherwise different trajectories before their differences have done useful work.

From model priors to authority-induced anchoring

A multi-agent system can be viewed as a pipeline with three interfaces. The model supplies its learned distribution and its tendencies toward or away from unusual continuations. The agent layer gives each participant a role, persona, or authority position. The system layer determines who sees which messages, in what order, and for how many rounds. A final editor or leader then turns the discussion into a concrete proposal. Diversity can be lost at any interface, and a later stage cannot always recover what an earlier stage removed.

classDiagram class ModelIntelligence { alignment intrinsic_variance } class AgentRole { persona authority } class CommunicationTopology { visibility speaking_order } class InteractionHistory { shared_context prior_judgments } class ConsensusPressure { agreement anchoring } class ProposalPool { semantic_diversity judged_quality } ModelIntelligence --> AgentRole : shapes AgentRole --> CommunicationTopology : operates within CommunicationTopology --> InteractionHistory : controls InteractionHistory --> ConsensusPressure : can increase ModelIntelligence --> ProposalPool : supplies priors ConsensusPressure --> ProposalPool : narrows or refines AgentRole --> ProposalPool : directs perspectives

At the model level, the multi-agent study finds a quality–diversity tension that is easy to misread. More capable or more strongly aligned models can produce fluent proposals and score well on correctness-oriented judgments while concentrating their semantic content. Increasing intrinsic variance can widen the accessible space, but it also makes quality less predictable. The source’s interpretation is that alignment can act as a global semantic regularizer: it narrows exploration without necessarily producing a comparable improvement in quality.

This matters for multi-agent design because an ensemble of agents built on a shared model does not automatically constitute a population of independent minds. Several agents can carry different labels while inheriting the same training distribution and alignment preferences. If the underlying model strongly favors a conventional research direction, the role prompts must overcome a common prior before they can create meaningful separation. Persona diversity may change the route into the model’s knowledge, but it does not erase the common model-level attractor.

The cognition layer can strengthen that attractor through authority. The study compares several collaboration structures, including flat groups of junior researchers, leader-led groups, interdisciplinary expert groups, vertical mixtures of senior and junior agents, and a less differentiated baseline.