Researchers have studied the traffic patterns generated by multi-agent systems that use large language models (LLMs). These systems create internal requests to the LLM backend, unlike traditional applications where user arrival rates govern request timing. The study analyzed three coordination topologies and found that topology significantly affects the arrival process of requests. For example, fan-out coordination introduces a bimodal distribution not seen in sequential execution. The researchers also found that a log-normal distribution best describes the reasoning phase component. Their framework and analysis pipeline are available online. --- Why it matters: This study matters to AI engineers because it highlights the need for new traffic models tailored to multi-agent systems, which could improve system performance and efficiency. Understanding how coordination topology affects request arrival patterns is crucial for designing scalable and reliable LLM-based applications. Source: https://arxiv.org/abs/2608.20494