NAIC 2026 · 3rd ACM SIGCOMM Workshop on Networks for AI Computing
Towards Traffic Modelling of Multi-Agent Systems: The Role of Coordination Topology
In agentic systems, where several LLM agents work on one task, a single request from a user can turn into dozens of calls to the model server, and the way the agents coordinate decides when those calls arrive. We ran the same tasks on the same model under three coordination patterns (agents taking turns, a star of reviewers, and everyone messaging everyone) and measured the traffic from the application down to the network. Fan-out turns the arrivals into bursts followed by long pauses, Poisson models don't fit, and two patterns with the same average call rate can load the server very differently.
This site has the code, a replay demo and the results. For the story in plain words, start with the blog post.
Tools
Run or replay the AgentVerse workflow, open a saved run, chat with the orchestrator, or browse the paper's results.
Run a workflow
Start an AgentVerse run, or replay a recorded one: recruitment → discussion → execution → evaluation, under the sequential, star or full-mesh topology.
Run a workflowOpen a saved run
Load a response.json or a run ID and step through its stages, discussion rounds and LLM calls.
Open a saved runCompare topologies
Replay the same task under the sequential, star and full-mesh topologies at once, with one shared timeline of LLM-call arrivals: the difference in burstiness at a glance.
Compare topologiesChat with Agent A
Send one task to the orchestrator and see its reply, the worker outputs and the LLM calls. Needs the live backend, so it works on a local install only.
Chat with Agent AResults
The paper's traffic findings from 1,481 recorded runs: inter-arrival times and fits, cross-layer correlations, and a drill-down into every run's LLM-call timeline.
Open Results