AgentAgora: Facilitated Multi-Agent Discussion for Creative Exploration

As large language models take on more open-ended tasks, creativity becomes increasingly important: design, inventive problem solving, and research ideation all require useful alternatives to familiar approaches. Multi-agent discussion offers a promising way to combine perspectives, but interaction can also draw agents toward the same directions. In particular, different initial roles do not ensure that later contributions address what remains unexplored. To address this gap, we introduce AgentAgora, a framework that coordinates exploration as discussion evolves. Agents with different perspectives contribute candidates to a shared record, while a chair maps the areas already covered and assigns complementary areas for the next round. As new contributions arrive, the chair redirects effort toward gaps and unresolved alternatives without proposing or rating candidates itself. This separation allows the same coordination mechanism to support idea generation, single-answer decisions, and research synthesis. Across these settings, AgentAgora improves evaluated output quality, ranking first under the four reported judges on LiveIdeaBench and improving CreativityBench answer accuracy by 2.44 percentage points over the strongest baseline. Ablations and discussion traces further support sustained, coverage-guided exploration. Together, these findings offer a practical route to creative multi-agent collaboration, with potential applications in design and discovery.