AI Agents Talking to Each Other Shake Up Internet Traffic, Study Finds

When you ask an AI assistant a question, it's often not a single computer answering. Increasingly, it's a team of AI programs working together — each one thinking, checking, and calling others for help. This team is called a multi-agent system, and these agents generate traffic in a very different way than humans do.

Researchers ran experiments with three layouts for agent teams: one after the other (sequential), a leader sending tasks out (star), and everyone talking to everyone (full mesh). They repeated each setup 500 times and watched the traffic. The key finding: the way agents are arranged changes the timing of their requests to the AI engine. When one agent fans out to many others, traffic becomes "two-peaked" — some calls come fast, others slow. That's very different from the smooth, random pattern that internet engineers usually expect.

This matters because the old math models used to design networks don't fit AI agents. Those models assume users act independently and at random. Agents instead send bursts and follow internal "thinking" steps, so traffic arrives in clumps. The researchers also found the "reasoning" phase — when the AI ponders — follows a pattern called log-normal, not the usual Poisson distribution used for phone calls or web clicks.

For everyday people, this could mean AI-powered tools become slower, more expensive, or less reliable unless networks are redesigned for this new reality. The team published its tools, so engineers can start adapting now.