This study investigates the use of communicative AI agents in designing mathematical tasks. It examines how a network of LLM (large language models) agents can facilitate mathematical task design through collaborative communication in a chat chain. Four specialized AI agents were instructed each focusing on a different perspective: mathematical content, linguistic sensitivity, competence orientation, and differentiation. The AI agents sequentially modified given mathematical tasks, with each contributing a unique focus to the task's evolution. The resulting tasks were evaluated by in-service teachers as human experts. This way, the qualitative study explores the potential of LLM agent networks in educational contexts. First findings suggest that AI agents can support teachers in the development of mathematical tasks for diverse learning needs, but at the same time require adaptation by teachers to the educational situation.