Researchers explore the effectiveness of scaling a team of LLM agents across various orchestration architectures, finding that team scaling's benefits are task-dependent and architecture-specific. They propose an exact generate-transform decomposition to explain these results, highlighting the importance of coverage and transformation in determining accuracy gains.