Researchers have conducted a study on how to decompose work among agents in large language models (LLMs) for tasks such as determining Value-Added Tax (VAT). They tested four different configurations with varying numbers of narrow agents, from one wide worker to five narrow ones. The results show that the intermediate configurations led to better accuracy, but the single agent did not outperform the orchestrated set. The study also found that the matched-token criterion was supported, suggesting that any advantage of the single agent could be explained by prompt-budget effects rather than agent-count effects. --- Why it matters: This research is important for AI engineers and researchers because it provides insights into how to design more efficient LLM-agent systems, which can have significant impacts on tasks such as VAT determination. The study's findings can inform the development of more effective and scalable architectures for these systems. Source: https://arxiv.org/abs/2608.23395