Large language models (LLMs) are shifting from single chatbots to collaborative agent teams. In such systems, agents critique, debate, and refine each other’s outputs to improve quality. Yet this process may come at a cost: reduced diversity. This paper presents a simulation study investigating a phenomenon we term Diversity Collapse, referring to the rapid reduction of semantic diversity observed in multi-agent LLM workflows. We tested three common workflows (single-agent self-refinement, generator–critic, and generator–critic–summarizer) across 20 open-ended tasks. The results indicate a consistent tendency: collaboration refines outputs but can also lead to early convergence toward narrow consensus, with lexical diversity dropping by up to 65%. Notably, this convergence occurs before any human interaction or selection takes place, creating what we call pre-interaction bias. We discuss implications for human decision-making and propose design strategies to preserve diverse AI perspectives.