Samenvatting

There is growing interest in agent-based debates that combine multiple Large Language Models (LLMs) to increase robustness. Still, their coordination dynamics and vulnerability to disinformation remain unclear. We present an experiment in which LLM-based agents position Dutch political parties along two ideological axes (the economic Left–Right and socio-cultural GAL–TAN dimensions) and exchange arguments over multiple rounds until reaching consensus or a predefined limit. Two factors are tested: (1) whether agents can observe one another’s positions (anchoring) and (2) whether one participant introduces persuasive but inaccurate arguments (disinformation). Hidden positions lead to longer, more argumentative exchanges, while visible positions produce shorter, convergence-oriented discussions. Introducing a disinformation agent prevents overall group consensus. Still, the collaborative agents’ final placements remain close to expert benchmarks. These findings suggest that LLM-based debates are shaped by alignment and pattern repetition rather than the simulation of argumentative deliberation in subjective domains.