"I celebrate grass": Parody and social division of labor in multi-agent collaborative poetry

AI is increasingly used in cultural productions, from writing quarterly reports to fanfiction. However, in producing such artifacts, AI often lacks normative competence: the ability to reason about and follow social norms for production and output. According to theories of normative reasoning, explicit invocation of normative commitments and social division of labor increase normative competence. Here, we study how single- and multi-agent turn-taking LLM designs influence explicit normative reasoning and division of labor among agents in a collaborative generation task. Using Japanese renga poetry as a testbed, we asked AI agents to complete a set of semantically linked verses in the voices of Edgar Allan Poe, John Keats, and Walt Whitman. Consistent with theory, we find that multi-agent turn-taking LLMs were more likely to invoke core linked-verse criteria in reasoning, relative to single agents, and tended to divide renga duties based on normative criteria. Turn-taking poems also had higher embedding similarity between stanzas, greater psychological (e.g., emotional, motivational) similarity between authors' voices within poems, and were judged by separate LLMs as having more coherence and higher-quality linking. At the same time, turn-taking poems had lower corpus-based author distinctiveness, less lexical overlap with the target authors, and more emphasis on structure. Both conditions produced poetry that was parodic in different ways. The findings may inform enterprise applications of vocal mimesis ("sound like me"), long-range creative benchmarks for artificial general intelligence, and commodity aesthetic production.