Towards Traffic Modelling of Multi-Agent Systems: The Role of Coordination Topology

In the digital world, networks are often thought of as pipelines carrying messages from one person to another. For decades, engineers have designed these systems based on a simple assumption: traffic arrives in a somewhat predictable rhythm, driven by human users clicking buttons, sending emails, or loading web pages. When a person waits a moment before sending the next message, the network sees a gap. When many people act at once, the network sees a rush. This human-driven pattern has been the foundation for planning how much capacity a network needs to handle the load.

However, a new kind of digital worker is emerging: the autonomous agent. These are software programs powered by large language models that can think, plan, and act on their own. Unlike a human who pauses to think before typing a reply, these agents can generate a complex chain of internal actions to solve a single problem. One user might ask a question, but the agent might decide to call a tool, ask a peer agent for advice, check a database, and then synthesize an answer. All of this happens automatically, creating a stream of requests that the network must carry. The question for network engineers is whether the old rules still apply when the traffic is no longer driven by human hands, but by the internal logic of these digital teams.

A team of researchers set out to measure exactly how these new systems behave. They focused on a specific aspect of network traffic: the time that passes between one request and the next. In the world of network engineering, this gap is a crucial clue. If the gaps are random and scattered, the system behaves in a way that is easy to predict. If the gaps cluster together or follow a specific pattern, the system behaves differently, potentially causing bottlenecks that standard designs cannot handle. The researchers wanted to see if the way these agents are organized—how they talk to each other—changes the rhythm of the traffic they generate.

To find the answer, the team built a controlled environment where they could run the same task hundreds of times with different team structures. They tested three distinct ways of organizing the agents. In the first setup, the agents worked one after another, like a relay race where each runner waits for the previous one to finish before starting. In the second setup, a central leader assigned tasks to several workers simultaneously, like a manager handing out assignments to a group of employees all at once. In the third setup, every agent could talk directly to every other agent at the same time, creating a dense web of communication. They ran each of these scenarios 500 times to gather a massive amount of data on how the requests flowed.

The results revealed that the way agents are organized fundamentally changes the nature of the traffic. When the agents worked in a simple line, one after another, the time between requests was relatively steady and predictable. However, when the researchers introduced parallel work, the pattern changed dramatically. In the setups where agents worked together at the same time, the traffic became "bimodal," meaning it had two distinct behaviors happening at once. One behavior was a sudden burst of activity where many requests arrived almost instantly, separated by tiny fractions of a second. This happened because the system sent out multiple commands at the exact same moment. The other behavior was a slower, more thoughtful rhythm that occurred while the agents were actually processing information and reasoning through their tasks.

The researchers discovered that the old idea that network traffic arrives randomly, like raindrops falling on a roof, does not fit these new systems. They tested a standard mathematical model that assumes requests arrive independently of one another and found that it was decisively wrong for all the agent setups they studied. Instead, the time it took for the agents to think and reason between requests followed a different pattern, one that is better described by a curve that accounts for how tasks can take varying amounts of time depending on complexity. This means that the internal logic of the software is creating a specific, structured rhythm that is very different from human behavior.

These findings matter because they show that the design of the agent team itself acts as a generator of traffic. If a system is designed to have many agents talk to each other at once, it will create sharp, intense bursts of network activity that can overwhelm servers, even if the total number of tasks remains the same. The researchers found that in the most connected setup, where every agent could talk to every other agent, the time between requests dropped to mere milliseconds for a large portion of the traffic, creating a surge that was nearly 600 percent faster than the sequential setup. This surge is not a glitch; it is a structural feature of how the system works.

The study also highlighted that these changes ripple down through the entire network. The way the agents coordinate affects not just the application layer, but also the underlying data connections and the load on the servers processing the requests. The researchers measured how many data packets were sent and how much memory was used, finding that the bursty nature of the parallel setups led to higher peaks in network usage. This suggests that simply counting the total number of requests is not enough to understand the load; engineers must also look at how those requests are spaced out in time.

Ultimately, this work provides the first clear picture of how the structure of a multi-agent system shapes the traffic it creates. It shows that as we move toward a future where software agents collaborate to solve problems, the networks supporting them will need to be designed with these new patterns in mind. The old models, built for human users, cannot accurately predict the behavior of these digital teams. By understanding that the coordination style of the agents creates specific, measurable patterns in the traffic, engineers can build systems that are robust enough to handle the unique rhythm of the autonomous future. The researchers have made their tools and data available to the public, inviting others to explore these patterns further as the technology continues to evolve.