Same task, same model, different traffic

Why agent traffic matters

Most traffic models start with people. Someone clicks, a request goes out, and arrival rates follow human behaviour. Agentic systems break that link. A single task from a user can turn into dozens of model calls, and their timing is set by how the agents coordinate, not by the user.

This is no longer a niche concern. Cloudflare reported in June 2026 that traffic from AI agents and other bots now surpasses that of people. There is a familiar lesson here too: in the 1990s, measurements of Ethernet, wide-area and web traffic showed that Poisson models badly underestimate how bursty real traffic is (Leland et al. 1994; Paxson and Floyd 1995; Crovella and Bestavros 1997). Each new kind of traffic has needed its own model. We wanted to know whether agents are next.

For people outside of networking, a traffic model describes when requests arrive and how much work each one brings, in a form simple enough to analyse. Operators use such models to size servers and links, to set rate limits and autoscaling rules, and to drive simulations and load tests. A model only helps if it matches how the traffic behaves.

The classic starting point is the Poisson process: requests arrive independently of each other at a constant average rate. The gaps between arrivals are then independent and exponentially distributed: the length of each gap does not depend on the previous ones, and the time left until the next request does not depend on how long it has been since the last one. A single number, the rate, describes everything. Its simplicity, easy queueing results and good fit for telephone calls made it the default, but it assumes arrivals are spread out evenly. If real traffic comes in bursts, queues grow longer and tail latency gets worse than a Poisson model predicts at the same average load. That gap is what this paper is about.

The experiment

Each run hands an AgentVerse workflow a task: a maths problem, a research question, a consulting brief or a small coding job. Four expert agents are recruited, discuss the task, then carry out and evaluate a plan. They all use the same small model, Llama 3.2 3B, served by a single vLLM instance.

The only thing we changed is how the agents discuss:

We ran each pattern about 500 times and logged every model call, alongside metrics from the model server, the containers and the network. Most of what follows comes from one quantity: the gap between the starts of consecutive model calls during the discussion¹.

Bursty arrivals

When agents take turns, every gap is a few seconds long, roughly the time the model takes to answer. Fan-out changes this. In star, 53% of the gaps are shorter than 50 ms, because the reviews are dispatched together. In full mesh the share is 38%, and for sequential it is 0%. The remaining gaps form a second, slower group set by how long the previous step takes to generate. The arrivals end up bimodal: a burst, then a wait.

The Poisson model does not fit

The usual first guess for arrivals is a Poisson process, which implies exponentially distributed gaps. For star and full mesh that fails almost by construction, since calls sent together cannot be Poisson. In sequential there are no bursts, and the exponential still fits worse than the alternatives (KS distance 0.40, against 0.03 for the log-normal).

Looking only at the slower gaps, those above 50 ms, the log-normal has the lowest AIC of the three distributions we tried in every pattern, though how well it fits varies: the KS distance is 0.03 for sequential but 0.08 and 0.09 for star and full mesh, so in the fan-out patterns the slow group has structure that a single distribution does not capture. For full mesh the Weibull is slightly closer by KS distance. The fits, tests and robustness checks are in the paper.

Same rate, different load

For us this is the most practical result. Star and sequential make discussion calls at almost the same average rate, about 0.21 per second. Yet star keeps more calls in flight at the model server: 1.6 on average against 1.0. Its calls also take longer, 11.6 s on average against 5.7 s, partly because each call handles more tokens, prompt and response together (2.7k per call against 1.8k). Full mesh finishes its discussion in 32 s against 81 s and averages nearly three calls in flight. Although it sends fewer bytes to the model server in total (39 kB per run against 51 kB), it does so in less than half the time, so its mean byte rate is 77% higher (about 1.1 kB/s against 0.6 kB/s).

An average rate hides this. A rate limiter or autoscaler sized on averages would treat star and sequential as the same workload, even though one arrives as a steady trickle and the other in bursts. It also means that burstiness in a trace from an agentic system may reflect its coordination structure, not a change in load.

Limitations

This is one framework, one small model, four kinds of task and a single machine. The topologies were fixed in advance and the agents did not call external tools. The exact numbers are specific to this testbed. We expect the pattern to carry over: fan-out creates bursts, and the coordination structure, not just the user, shapes what the server sees.

Future work

Next we want more models and frameworks, deployments spread over several machines, and a mixture model that describes both groups of gaps. We also plan a semi-Markov model of the agents' execution states, with transition probabilities estimated from our traces, that can generate synthetic agent traffic.

Try it

The measurement framework and analysis scripts are on GitHub. In the browser you can replay a full-mesh run and watch the edges of the topology diagram light up as calls land, or explore the results in more depth. The slides from the NAIC talk are here:

• NAIC 2026 talkThe workshop presentation of the paper Checking…

Citation

If this is useful for your work, please cite the paper:

D. Lamagna, A. Cabellos, A. Rodriguez-Natal, G. Rétvári and B. Serracanta. Towards Traffic Modelling of Multi-Agent Systems: The Role of Coordination Topology. In 3rd ACM SIGCOMM Workshop on Networks for AI Computing (NAIC). ACM, 2026. DOI .