Source-linked AI summary
Towards Traffic Modelling of Multi-Agent Systems: The Role of Coordination Topology
Davide Lamagna, Albert Cabellos, Alberto Rodriguez-Natal, Gábor Rétvári, Berta Serracanta
TL;DR
Multi-agent LLM systems generate internal traffic whose timing depends on coordination logic, raising questions about whether classical human-driven traffic models apply. The paper empirically measures LLM-call inter-arrival distributions across three topologies using a multi-layer framework, finding topology-shaped arrival processes, fan-out bimodality, and log-normal reasoning-phase behavior with exponential arrivals rejected.
Problem
Multi-agent systems generate internally coordinated LLM requests, but evidence is limited on how coordination topology shapes their arrival processes and whether classical traffic models apply.
Method
The paper measures LLM-call inter-arrival times across sequential, star, and full-mesh topologies using a multi-layer framework spanning application, model-serving, container, and network layers.
Results
Topology determines whether arrivals are homogeneous or bimodal with a fan-out component; log-normal best fits the reasoning-phase component, while the exponential model is decisively rejected across all topologies.
Takeaways & Limitations
Coordination topology acts as part of the workload generator, and its structural differences propagate to inference and network-level behavior.
Takeaways & Limitations
The study focuses on LLM-call traffic, leaving external tool calls and actions for future work; broader model and serving configurations are also proposed for study.
Abstract
from arXiv · showhide
Multi-agent LLM systems are an emerging networked workload whose rapid deployment raises questions about the traffic patterns they generate. Compared to conventional applications, these systems generate requests internally: a single user task can induce a structured sequence of model calls whose timing is governed by coordination logic rather than by user arrival rate. It is not clear whether classical traffic models, designed for human-driven workloads, apply to this setting. We present an empirical characterisation of LLM-call interarrival time distributions across sequential, star, and full-mesh agentic coordination topologies, using a multi-layer measurement framework over 500 repeated runs per topology. We find that topology fundamentally shapes the arrival process of requests to the LLM backend: fan-out coordination introduces a structural bimodality absent in sequential execution, and the reasoningphase component is best described by a log-normal distribution, with the Poisson exponential null model decisively rejected across all topologies. These differences propagate to inference and network level metrics. The framework and analysis pipeline are released openly at https://github.com/dlamagna/agentraffic.
1 Introduction
Multi-agent systems generate internally timed LLM calls, making coordination topology a potential determinant of traffic patterns. This paper measures inter-arrival processes across sequential, star, and full-mesh topologies and introduces an open measurement framework.
- Motivation: A single user task can induce internal LLM calls and inter-agent messages whose timing is shaped by coordination logic rather than user arrival rate.Internal branching, retries, and information exchange can create bursts affecting shared inference and networking infrastructure.
- Motivation: Traffic characterisation is motivated by prior evidence that new workload paradigms can invalidate classical assumptions such as Poisson arrivals.The paper positions agentic systems within an empirical tradition of measuring emerging network workloads.
- Research question: The paper asks whether coordination topology induces distinct request-arrival processes and uses LLM-call inter-arrival times as its primary traffic metric.TCP and packet metrics provide lower-layer context for the request-level measurements.
- Study design: The study compares sequential, star, and full-mesh topologies, representing ordered execution, orchestrator-driven fan-out, and dense peer-to-peer exchange.The goal is to measure how coordination structure appears in the arrival process rather than exhaustively cover the design space.
- Contributions: Distribution fitting finds log-normal to be the best model across topologies, while decisively rejecting the exponential null for reasoning-phase inter-arrival times.The paper also releases an open-source, multi-layer framework for configurable topology experiments.
2 Background
LLM agents generate internally coordinated traffic whose timing can differ from conventional user-driven request streams. Inter-arrival times and coordination topology provide key ways to characterize these workloads and their effects on traffic modeling.
- LLM agents as traffic-generating systems: LLM agents can generate additional model work during a task, so observed request arrivals may reflect internal control flow rather than user arrival rate.ReAct-style agents decide whether to query models, use tools or agents, or continue execution.
- LLM agents as traffic-generating systems: A single external request can expand into a dynamically shaped internal call graph governed by reasoning, coordination policy, and model or tool outputs.The paper focuses on LLM-call traffic, leaving external tool calls and actions for future work.
- LLM call latency and inter-arrival time: Inter-arrival time is the wall-clock gap between consecutive requests and helps assess traffic-model compatibility, queuing behaviour, and resource dimensioning.It is considered alongside burstiness, correlation structure, and flow or request sizes.
- LLM call latency and inter-arrival time: Parallel dispatch can create near-zero inter-arrival gaps between concurrently issued long-running calls, unlike sequential workflows where IAT is strongly coupled to call latency.Phase-structured coordination may produce distinct timing patterns across recruitment, dispatch, discussion, aggregation, and finalization.
- Coordination topology: Multi-agent frameworks expose coordination topology as an implementation choice through programmable conversations, staged phases, graph-structured workflows, and role-specialised orchestration.Examples include AutoGen, AgentVerse, LangGraph, and CrewAI.
3 Measurement framework
The measurement framework treats coordination topology as the controlled design variable while observing agentic traffic across application, model-serving, container, and network layers. It combines passive infrastructure measurements with application traces and shared identifiers and timestamps to connect request-level topology effects across layers.
- The framework measures agent coordination across application, model-serving, container, and network layers.Topology is varied while the agent runtime, model backend, and monitoring stack remain unchanged.
- The testbed uses containerised LLM, orchestrator, and worker services connected through Docker bridge networks.Packet capture reconstructs TCP flows, while Jaeger and cAdvisor provide application-trace and container-resource measurements.
- AgentVerse experiments compare sequential, star, and full-mesh discussion coordination with four recruited subagents.Sequential calls agents one at a time, star dispatches parallel reviewers, and full mesh submits directed peer-message pairs concurrently.
- Metrics are collected primarily outside the agent execution path, with network and container measurements observed passively and application traces supplying semantic context.The framework organises these measurements by architecture layer.
- Shared agent identifiers and timestamps allow per-run metrics to be joined and correlated directly across layers.Request-level inter-arrival times are the primary workload model, while TCP, connection, and packet metrics provide cross-layer evidence.
4 Experiment design and results
Across 500 repeated runs per topology, coordination structure produced distinct LLM-call arrival patterns, with fan-out topologies creating bursty bimodal traffic. Reasoning-phase gaps were evaluated with distribution fitting, where log-normal generally provided the strongest overall fit.
- Experiment design: 500 repeated runs compared sequential, star, and full-mesh coordination using IAT as the primary traffic metric.All runs used the same task source, agent implementation, inference backend, and bridge-level instrumentation.
- Cross-layer comparison: Topology changes were visible across application, inference, and network layers, including median IAT, concurrency, token throughput, and TCP byte rate.Per-run correlations linked token throughput, in-flight requests, and TCP byte volume, while larger IATs were negatively associated with backend traffic rate.
- Topology effects on IAT: 6,513 sequential IAT samples had mean 6.23 s, median 4.56 s, and p95 10.79 s, forming a homogeneous multi-second arrival process.Sequential coordination served as the baseline, with agents contributing in strict order.
- Topology effects on IAT: Star produced 7,355 samples with mean 6.88 s, median 0.41 ms, and p95 20.74 s; 53.3% of gaps were below 50 ms.Full mesh produced 14,493 samples with mean 1.09 s, median 236 ms, p95 2.82 s, and 38.2% below 50 ms.
- Bimodality and distribution shape: Star and full mesh showed structural bimodality caused by near-simultaneous subagent fan-out, unlike approximately unimodal sequential traffic.The fan-out mode is defined as IAT ≤50 ms, while the reasoning phase is IAT >50 ms.
- Goodness-of-fit analysis: Log-normal achieved the lowest AIC for all three topologies, while Weibull had a lower KS statistic than log-normal for full mesh.For full mesh, Weibull’s KS statistic was 0.068 versus 0.089, while log-normal’s AIC was 20,590 versus Weibull’s 21,760.
- Goodness-of-fit analysis: Subsampling robustness checks found that log-normal was rejected least frequently, supporting a non-memoryless reasoning-phase arrival process.The fitting analysis restricted model comparison to IATs greater than 50 ms and evaluated Exponential, Weibull, and Log-normal families.
5 Discussion
The coordination graph acts as part of the workload generator, changing whether agents wait, fan out, or exchange messages concurrently. These topology-dependent arrival structures affect modelling needs and extend to inference and network behaviour.
- Topology changes the LLM-call arrival process, so identical average task rates can impose different backend demands.Serial execution and coordinated fan-out generate different temporal load shapes.
- Reasoning-phase gaps are non-memoryless: exponential fitting is decisively rejected across all topologies, while log-normal fits best.Call latency can reflect prompt length, context accumulation, output token count, and batching decisions acting multiplicatively.
- Star and full-mesh traffic mixes reasoning-phase arrivals with structural fan-out, preventing one parametric family from describing the full distribution.This produces bimodal arrival distributions in those topologies.
- Burst-sensitive summaries such as IAT percentiles, sub-second burst fractions, concurrency peaks, and mixing fractions complement aggregate throughput.These measures capture differences that average task rates can conceal.
- Topology adds temporal load on shared infrastructure alongside solution quality, robustness, and token cost as a system-design consideration.The paper frames traffic shape as an additional dimension for multi-agent design.
- The reported quantitative values are controlled-testbed measurements from one task family, framework, model, single-host deployment, and bridge-level capture point.The specific table parameters are not intended to generalise beyond this configuration.
6 Conclusion and future work
Across sequential, star, and full-mesh coordination, topology determines whether arrivals are homogeneous or bimodal, while log-normal fits the reasoning-phase component and exponential is rejected. Future work broadens the workload matrix and develops a generative model, supported by an open-source framework.
- 6 Conclusion: Topology determines whether LLM-call arrivals form a homogeneous reasoning-phase distribution or a bimodal mixture with fan-out.This pattern spans sequential, star, and full-mesh coordination.
- 6 Conclusion: Log-normal is the best fit for the reasoning-phase component under all three topologies, while exponential is decisively rejected.No single distribution captures the full arrival process.
- 6 Conclusion: Cross-layer measurements show that topology-dependent arrival structures affect inference and network-level behaviour beyond the application layer.
- Future work: Future experiments should vary model size, model family, serving configuration, and external tool or data-service calls.These additions would test broader operating ranges and introduce arrival classes beyond LLM calls.
- Future work: A semi-Markov model of agent execution states could use trace-estimated transition probabilities to build a generative arrival-process model.
- Future work: The open-source framework is intended to measure agentic traffic across different agent stacks, model backends, and deployment topologies.