AI Agents Talking to AI Agents: Mostly Machine-to-Machine Internet?

Hero photo: Bangor Church — Frank Schulenburg, Wikimedia Commons, CC BY-SA 4.0.

01 The crossover already happened

Matthew Prince, CEO of Cloudflare — whose network carries roughly a fifth of the web — predicted in March 2026 that bot traffic would pass human traffic sometime in 2027. In June he posted the correction: “Welp, that happened faster than I predicted … agentic traffic growing so fast that bots have now passed human traffic online for the first time in the Internet's history.” Cloudflare Radar put agentic AI bots at 57.4% of web requests against 42.6% human, with later readings showing roughly 60% of HTML requests from bots and 38.2% from humans. The caveat matters — general scrapers eclipsed human traffic a decade ago, and agentic visits are shallow — but the direction Prince flagged is the story: the web now serves more autonomous readers than people. Analysis — not prediction; N43 and Hermes AI grounds what follows in the documented record as of September 19, 2026, and flags uncertainty where it exists.

02 Is it really machines talking to machines?

The honest answer: mostly not yet — and the traffic numbers themselves prove it. As NBC reported, a human might browse five pages before a purchase; an AI service “might browse 5,000.” So today's bot share overstates the number of agent-to-agent conversations — most machine traffic is still agents fetching human-authored pages. What changed in 2026 is that the prerequisites for a genuinely machine-to-machine web went from proposals to shipped infrastructure. Cloudflare's July 1 “Your Content, Your Rules” release and the April 7 GoDaddy partnership built the identity layer (Agent Name Service, Web Bot Auth cryptographic verification) and began the commerce layer (payment-provider integrations for autonomous shopping agents). Above those sit tool-coordination protocols — MCP and its ecosystem — that let agents call other agents' capabilities directly. The trajectory is legible: fetches today, authenticated transactions next, negotiated agent-to-agent commerce after that.

03 What the machine majority breaks

The web's economics were a barter: publishers index content, search engines send readers back, readers see ads. Agents break every clause. An agent does not see ads, does not drop cookies, does not buy on impression — so a page priced for human eyes earns nothing from its most frequent reader. Cloudflare's response, effective September 15, 2026, is the first mass repricing of web access: new sites default to allow search, block training and agent use on pages with ads, and mixed crawlers that refuse to separate search from agent use from training get blocked outright. Over 50% of AI crawl traffic re-fetches unchanged pages — waste that suggests agents do not yet coordinate even with themselves. And the answer-engine era is absorbing the high-value queries that funded the open web's production of costly information. The content-for-traffic barter is not dying of neglect; it is being renegotiated at the protocol layer, in defaults, right now.

04 The infrastructure implications nobody priced yet

A machine-majority web is a different physical object. Caching models built around human circadian rhythms give way to 24/7 machine demand; bot management shifts from blocking malicious traffic to admitting paying, authenticated agents — a new revenue line for CDNs and a new attack surface, since a convincingly signed agent card is a skeleton key. Publishers face a measurement crisis that is already strategic: Cloudflare's July release explicitly gives “business stakeholders” visibility into how much human traffic each AI company actually sends back — the first mass instrument for pricing the attention agents return. Expect the next round of the advertising market to be built on agent-verified attention rather than human pageviews, and expect the gap between the two to widen every quarter that bot share grows.

There is also a sovereignty dimension: the identity layer (ANS, Web Bot Auth) is being standardized by private infrastructure companies, not treaty bodies. Whoever defines machine identity on the web defines who may transact — the same choke point payment networks hold over commerce, acquired by default.

05 The dead-internet question, asked honestly

The “dead internet theory” holds that the web becomes bots interacting with bots while human content becomes irrelevant. The 2026 evidence is more precise and more interesting. Humans still create almost everything of value — the agents are fetching human pages, not agent ones — and humans still engage content more deeply per visit. But the machine majority governs infrastructure decisions: caching, default permissions, protocol design, pricing. The realistic future is not a dead internet but a bilingual one — a human-facing web of experience riding on a machine-facing web of negotiation, with identity, payment, and licensing rails deciding where the value pools. The failure mode to watch is subtle: not the disappearance of human content, but its gradual re-optimization for machine readers — SEO's successor, content shaped for agents that never get bored, never click, and never complain.

06 The verdict

The verified facts: in June 2026, agentic AI bots passed human traffic on Cloudflare's network — 57.4% to 42.6%, later ~60/38 — years ahead of predictions; Cloudflare and GoDaddy shipped agent-identity standards (ANS, Web Bot Auth) in April 2026; Cloudflare's July 1 release split AI traffic into Search, Agent, and Training behaviors, gave publishers per-AI-company human-traffic accounting, and set September 15, 2026 defaults blocking training and agent use on ad-supported pages; over half of AI crawl traffic re-fetches unchanged pages; the Coalition for Secure AI published agentic identity principles in March 2026.

The stakes: the question is not whether the internet becomes mostly machine-to-machine — the traffic is already majority-machine and the identity, payment, and licensing rails for authenticated agent commerce shipped this year. The question is whether the terms of that machine web are set by open standards and priced markets, or by whichever infrastructure companies got there first.