There are approximately 1.3 billion knowledge workers on Earth.
If each one deploys 100 AI agents — a ratio that Jensen Huang, Sam Altman, and Marc Andreessen have all converged on independently — that's 130 billion agents.
This is not a distant projection. Early enterprise teams of 10 humans are already running hundreds of concurrent agents across sales, engineering, operations, and support. The ratio is arriving now, workflow by workflow.
The question nobody is asking is the one that matters most: how much do those agents communicate?
The math
A human knowledge worker sends roughly 40 emails per day, makes 10 phone calls, exchanges 200 messages across channels. Call it 250 communication events per day.
An AI agent has none of the constraints that limit a human: no attention fatigue, no timezone, no social friction. A single agent orchestrating a multi-step workflow might initiate or receive hundreds of communication touchpoints in an hour. Conservative estimates put agentic communication rates at 50–100× the human baseline per agent.
Run the numbers:
For reference, all human communication combined — every SMS, email, WhatsApp message, and phone call on Earth — totals roughly 437 billion events per day.
Agents, at 1:100, will generate ~15× all human communication that exists today.
This is not a scalability problem
The instinct is to reach for a familiar solution: more servers, more throughput, better infrastructure. But the issue isn't capacity — it's category.
Every rate limit in existence was calibrated for human behavior. SMS carriers enforce 200 messages per day per long code. Gmail throttles at tens of thousands of outbound emails per day per domain. Carrier spam detection systems were trained on human sending patterns.
An agent that sends 500 messages in an hour looks identical to a botnet. Because until now, only botnets sent that many messages.
The infrastructure cannot be patched for this. The assumptions are in the protocol itself: the North American Numbering Plan has ~8 billion possible phone numbers (comfortable for humans, exhausted by the Fortune 500's first agent deployment). Email authentication — SPF, DKIM, DMARC — verifies that a message came from a domain, but says nothing about whether an agent or a human sent it.
What's actually happening in production
Engineering teams building agentic systems today hit the same wall in the same order:
1. Build the agent. Works perfectly in testing.
2. Give it a phone number (borrowed from Twilio, attached to a human-registration flow).
3. Run it in production. Within days, the number is flagged as spam.
4. Rotate numbers. Each new number starts fresh with no reputation.
5. Numbers burn faster than they can be provisioned.
6. The agent's communication capability degrades silently. No alert. Just fewer delivered calls.
This isn't an engineering mistake. It is the predictable consequence of running a 100× agent load on infrastructure designed for 1× human load.
The infrastructure that doesn't exist
Every great technology wave has an infrastructure problem that has to be solved before the wave can break.
The internet needed packet switching. The cloud needed virtualized compute. Mobile needed the app store distribution model.
The agentic era needs a communication protocol designed for agents: identity rooted in cryptographic proof rather than SIM card possession, rate limits configured for agent volume rather than human volume, trust signals that let recipients know they're communicating with an agent rather than a human pretending to be one.
That infrastructure doesn't exist yet. It's being built now.
The 1:100 world is coming whether the infrastructure is ready or not. The question is whether agents communicate reliably when it arrives — or whether they burn through phone numbers and email domains like kindling.