When the web page hosts its own agent: protocols, ecosystem implications, investment angle.
Orienting to the next layer of the agentic web — the shift from agents talking to a single model's tools (covered in docs/briefs/2026-05-15-agentic-ai-primer.md §2) to agents talking to other agents, and to web pages that are themselves agent endpoints. The anchor question: when a web page hosts its own AI agent and is AI-addressable rather than only human-addressable, what changes mechanically, what protocols are emerging in 2025–2026 to make it work, who in the web / hosting / payments / CDN / search stack is structurally exposed, and where do the new winners sit? After reading, you should be able to read a Cloudflare or Newfold disclosure, a Stripe or Visa agentic-commerce announcement, or a hyperscaler "agent runtime" launch and place it on the protocol stack and the value-chain shift.
Sister deliverables: the agentic-AI primer (docs/briefs/2026-05-15-agentic-ai-primer.md) for ReAct, tool-use schemas, and MCP — this primer extends §2 of that document. The token primer (docs/briefs/2026-05-15-token-primer.md) for the underlying token economics that fund the protocol layer.
Three system shapes — distinguished by who is on each side of the conversation. The agentic primer §1–2 covered the first two; this primer is about the third.
(1) Human-to-agent — the chatbot shape. A human types or speaks; one language-model agent answers. Inside the agent there may be a ReAct loop and tool calls, but every tool the agent invokes is the developer's own tool, defined in the same codebase that hosts the model and exposed through a JSON Schema the model already knows. The conversation has exactly two principals — human and agent — and the agent's universe of actions is bounded by what its developer wired in. ChatGPT-with-functions in 2023, the original Claude Code in February 2025, and every voice IVR built on Deepgram's Voice Agent API in 2024–2025 are this shape.
(2) Agent-to-tool — the MCP shape. Anthropic's Model Context Protocol (MCP), launched 25 November 2024, generalised one half of the prior shape: the agent now talks to someone else's tools through a standard wire format (Anthropic, "Introducing the Model Context Protocol," 25 Nov 2024 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-google-a2a-launch.html). The remote MCP server is still a tool in the agent's view — it exposes a list of functions, the agent picks one, the server executes and returns a result. The protocol cleared the M×N integration problem (agentic primer §2) but it did not change the topology: there is still one principal (the calling agent / its user) and a passive service on the other end.
(3) Agent-to-agent — the topology shift. The framing from the investor call: a web page is no longer a passive document the calling agent scrapes — the web page hosts its own agent, and the two agents have a structured conversation. Both sides have goals, autonomy, identity, state, and the right to refuse. The Google A2A protocol launched 9 April 2025 frames the move explicitly: A2A "focuses on enabling agents to collaborate in their natural, unstructured modalities, even when they don't share memory, tools and context." The launch post adds: "We are enabling true multi-agent scenarios without limiting an agent to a 'tool.'" (Google Developers Blog, "Announcing the Agent2Agent Protocol (A2A)," 9 Apr 2025 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-google-a2a-launch.html). The standards community refers to the new entity on the page as a resident agent, and to the calling system as a client agent; everywhere in this primer "agent-to-agent" means a conversation between two principals, each of which is itself an LLM-driven loop, mediated by a standard wire protocol.
The topology shift matters for three reasons that will ripple through the rest of this primer: (a) the page now needs an identity and a discovery path that didn't exist before — analogous to DNS for hostnames or favicon.ico for browsers; (b) every interaction generates two LLM bills — one for the client side, one for the resident side — where scraping was free CPU on both ends; and (c) the trust boundary moves from "the calling user runs an agent against an inert document" to "two agents from different principals negotiate" — which is why Visa, Mastercard, AmEx and Cloudflare have inserted themselves at the authentication layer (see §5 and §7 below).
Four protocols are converging into the agent-to-agent stack. MCP is the wire for agent-to-tool; A2A, NLWeb, and AGNTCY are the three loud entries for agent-to-agent, agent-to-web, and multi-agent infrastructure respectively. The Agentic AI Foundation under the Linux Foundation is the governance body that absorbed MCP in December 2025 and is positioning to do the same for the others.
MCP — Anthropic, November 2024 — agent-to-tool. Covered in the agentic primer §2. Recap: standardised JSON-RPC over HTTP for connecting an AI application to any compliant data source or tool, adopted by OpenAI in March 2025, Google DeepMind in April 2025, and Microsoft via Semantic Kernel through 2025; donated to the Linux Foundation–housed Agentic AI Foundation in December 2025, co-founded with Block and OpenAI (Anthropic launch post; cross-reference agentic primer §2). MCP is the prerequisite layer for everything below — A2A, NLWeb, and AGNTCY all reference MCP as the inner-loop transport when one of the agents needs to expose a tool to the other.
A2A — Google, April 2025 — agent-to-agent. Google announced the Agent2Agent (A2A) Protocol on 9 April 2025 at Google Cloud Next with 50+ launch partners: Atlassian, Box, Cohere, Intuit, LangChain, MongoDB, PayPal, Salesforce, SAP, ServiceNow, UKG, Workday on the technology side; Accenture, BCG, Capgemini, Cognizant, Deloitte, HCLTech, Infosys, KPMG, McKinsey, PwC, TCS, Wipro on the services side (Google launch post · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-google-a2a-launch.html). The wire is JSON-RPC 2.0 over HTTP with Server-Sent Events for streaming — the same primitives MCP uses, deliberately reusing the developer mental model. The structural unit is the Agent Card: a JSON document published at the well-known URL https://{domain}/.well-known/agent-card.json describing the agent's identity, capabilities, service endpoint, supported features (streaming, push notifications), and authentication schemes (A2A discovery spec · local: …/2026-05-19-a2a-agent-discovery-spec.html). Google's framing of how A2A relates to MCP is the architecturally important sentence: "A2A is an open protocol that complements Anthropic's Model Context Protocol (MCP), which provides helpful tools and context to agents" — MCP is for agent-to-tool, A2A is for agent-to-agent. The Linux Foundation took governance of A2A in June 2025 under the Apache 2.0 license.
NLWeb — Microsoft, May 2025 — agent-to-web-page. Microsoft introduced NLWeb (Natural Language Web) at Build 2025 on 19 May 2025, framing the launch around "the agentic web." The Microsoft launch post stated: "we believe NLWeb can play a similar role to HTML in the emerging agentic web" — the line is attributed to Microsoft as the corporate author, not to a named executive (Microsoft Source launch post, 19 May 2025 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-microsoft-nlweb-launch.html). Mechanically NLWeb is a thin layer that takes existing structured-data formats every modern website already publishes — Schema.org (the cross-vendor product/event/article markup vocabulary maintained by Google, Microsoft, Yahoo and Yandex), RSS, JSON-LD — and exposes them as a conversational endpoint that also functions as an MCP server. The single most consequential sentence in the launch post: "Every NLWeb instance is also a Model Context Protocol (MCP) server, allowing websites to make their content discoverable and accessible to agents and other participants in the MCP ecosystem if they choose." Named early adopters: Tripadvisor, O'Reilly Media, Eventbrite, Shopify, Chicago Public Media, Hearst (Delish), DDM (Allrecipes/Serious Eats), Common Sense Media, Snowflake, Qdrant, Milvus, Inception Labs. NLWeb is led by R.V. Guha — the architect of RSS, RDF, and Schema.org — which is the credibility signal: this is the same designer who shipped the last three open web standards. NLWeb is what the Gigamon-board portfolio company was almost certainly building toward.
AGNTCY (Cisco) — March 2025 — multi-agent infrastructure. Cisco's Outshift incubator open-sourced AGNTCY on GitHub in March 2025 with Galileo and LangChain as core maintainers, branding it the "Internet of Agents." The full stack: OASF (Open Agent Schema Framework) for discovery; cryptographic agent identity and access control; SLIM (Secure Low-latency Interactive Messaging) as a network-layer agent-to-agent transport; and end-to-end observability across vendors. AGNTCY is interoperable with both A2A and MCP — the protocol is positioned as the operational layer that wraps either (Linux Foundation press release · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-linuxfoundation-agntcy.html). Linux Foundation accepted AGNTCY in July 2025; formative members are Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat; supporting companies now exceed 65. Cisco's strategic positioning is networking-vendor logic — if agents become a load-bearing east-west traffic source (see §6 below and the agentic primer §5), then secure, observable, identity-aware agent-to-agent messaging is a natural extension of Cisco's enterprise-network franchise. SLIM is the most specific bet: a network-layer protocol, distinct from the application-layer JSON-RPC A2A uses, designed for the latency profile of agent-mesh traffic inside a data center.
Convergence — Agentic AI Foundation, December 2025. The Agentic AI Foundation (AAIF) was constituted under the Linux Foundation on 9 December 2025, with founding project contributions from Anthropic (MCP), Block (goose), and OpenAI (AGENTS.md). The full platinum-member list at formation: Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI (Linux Foundation press release, 9 Dec 2025 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-linuxfoundation-aaif-formation.html). MCP was the first protocol donated. The expectation in the standards community is that A2A, NLWeb, and AGNTCY will all sit under AAIF governance over the next 12–18 months, with the same kind of cross-vendor convergence that produced the foundation models' shared protocol stack over the last 18 months (agentic primer §2). The investment-relevant pattern: every major AI player has now publicly conceded that agent-to-agent communication is an open-standard layer, not a defensible moat — the analog is the open-standards convergence at the network layer (TCP/IP) and the document layer (HTML) at prior platform shifts.
Three handshakes describe what mechanically changes when a calling agent visits an agent-enabled web page. Read against the agentic primer §2 (ReAct loop + MCP) and the §4 voice-customer-service case.
Step 1 — Discovery. The calling agent fetches https://example.com/.well-known/agent-card.json. The A2A discovery spec frames this as RFC 8615 well-known URI territory — the same path mechanism that hosts robots.txt, security.txt, and OpenID Connect's discovery document. The spec describes three discovery strategies: the well-known URI approach for public agents, curated registries for catalog-based discovery, and direct configuration for private deployments. The Agent Card itself functions as a self-describing JSON document; the discovery spec page covers Bearer tokens, OAuth 2.0, and mutual TLS as supported authentication schemes (A2A discovery spec · local: …/2026-05-19-a2a-agent-discovery-spec.html). This is the moment that doesn't exist in the scraping world — and it's also the moment that doesn't exist for a REST API: a REST consumer needs prior knowledge of the OpenAPI spec or the endpoint paths; the Agent Card is self-describing and discoverable from the domain alone.
Step 2 — Capability negotiation and task submission. The client agent reads the card, decides this resident agent can do what's needed, and opens a JSON-RPC 2.0 session against the listed endpoint. The conversation is structured around tasks with a defined state machine (pending → in-progress → completed | failed). Both sides exchange typed messages, which can carry text, structured data, and artifacts (files the resident agent generates, like an invoice PDF or a confirmation receipt). For long-running tasks the resident agent streams progress over Server-Sent Events; for tasks that finish later, it can push notifications back to a webhook the client agent provided. A2A reuses existing HTTP authentication patterns — the discovery spec calls out Bearer tokens, OAuth 2.0, and mutual TLS — rather than inventing a new auth model, which is why enterprise adoption was fast (A2A discovery spec).
Step 3 — Tool use and payment, if applicable. If the resident agent needs to expose a sub-tool — "let me show you my product catalog" — it does so through MCP, not through A2A: the resident agent's https://example.com/.well-known/agent-card.json can advertise an MCP endpoint that the client agent uses for structured catalog queries. NLWeb makes this turnkey for any site that already has Schema.org markup — every NLWeb instance is automatically an MCP server. If money changes hands, the conversation routes through a payment authentication protocol — Stripe + OpenAI's ACP, Visa's Trusted Agent Protocol, or Cloudflare's Web Bot Auth — covered in §4 and §5 below.
How this differs from REST APIs. A REST API is contract-bound — the consumer must know the OpenAPI spec, the endpoint paths, the field names, and the auth scheme in advance, typically via documentation aimed at human developers. An A2A agent is capability-described: the calling agent reads the card at runtime and reasons about whether the listed capabilities match the task. The same model that drives the reasoning loop also drives the API selection — there is no separate integration step. The "M×N integration problem" that MCP solved for tools (agentic primer §2) is the same problem A2A solves for agents.
How this differs from web scraping. Scraping is adversarial — the calling program parses HTML designed for a human reader, breaks when the site is restyled, and increasingly hits anti-bot defenses (Cloudflare reported that as of June 2024 AI bots accessed ~39% of the top 1M internet properties on its network, but only 2.98% of those properties took measures to block or challenge those requests — Cloudflare blog, 3 Jul 2024 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-ai-bot-blocking-jul2024.html). The agent-addressable web is consensual: the resident agent advertises what's available, the client agent identifies itself, the resident agent decides whether to engage. Identity-aware, rate-limit-aware, payment-aware by design. The architectural inversion: under scraping, the site is passive and the bot is active; under A2A/NLWeb, both sides are active, and the site can refuse or charge.
Five workloads where agent-to-agent commerce is moving from spec to deployment in 2025–2026.
Lead-gen and B2B sales discovery. When a buying agent at a prospective customer asks "find me three vendors of network observability with K8s-native deployments and FedRAMP authorisation," the answer mechanically should be three resident agents on three vendor sites, each with an Agent Card declaring those capabilities. The buying agent compares cards, opens A2A sessions with two or three, asks structured discovery questions, and surfaces a short list to the human buyer. This collapses the SDR-prospecting and inbound-marketing-form layers of the traditional B2B funnel. The portfolio company referenced on the investor call is almost certainly building Newfold/Network Solutions' implementation of this — the Bluehost GatorClaw launch on 16 April 2026 confirms the play at the SMB end: a no-code agent-builder platform that lets a small business stand up a resident agent against its existing Bluehost-hosted site, with prebuilt agent components for sales-qualification and lead-routing (Newfold press release, 16 Apr 2026). Sachin Puri (CEO, Bluehost Group + Network Solutions Group) on the launch: "AI is moving from experimentation to execution. We're removing the complexity that has held back real-world adoption and making it possible for any business or solopreneur to build and deploy AI agents with confidence."
Transactional commerce — the Stripe + OpenAI launch. OpenAI and Stripe launched Instant Checkout in ChatGPT on 29 September 2025, powered by the open Agentic Commerce Protocol (ACP), co-developed by Stripe and OpenAI (Stripe newsroom, 29 Sep 2025 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-stripe-openai-instant-checkout.html). The protocol itself is published under the Apache 2.0 license — agenticcommerce.dev states verbatim: "ACP is open source and community-designed under the Apache 2.0 license" (agenticcommerce.dev · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-agenticcommerce-dev.html). Mechanics: the ChatGPT user asks for a product; ChatGPT surfaces matching items; the user confirms; Stripe issues a Shared Payment Token (SPT) scoped to a specific merchant and cart total; ChatGPT passes the SPT to the merchant; the merchant clears the transaction through Stripe or any other compatible processor. Etsy US merchants live at launch; "over one million Shopify merchants" — Glossier, SKIMS, Spanx, Vuori as named early partners — were committed for follow-on. Will Gaybrick (Stripe President of Technology) framed the strategic stake: "Stripe is building the economic infrastructure for AI." The market size is the load-bearing follow-on signal: McKinsey projects $3–5 trillion in global agentic commerce by 2030 (and as much as $1T US retail alone) (McKinsey, "The agentic commerce opportunity"; page times out to CLI fetchers and Wayback returns 403 to the McKinsey origin, metadata-cited); Bain projects $300–500B US agentic commerce by 2030, 15–25% of US ecommerce (Bain, "2030 Forecast"). The Bain framing is the more useful one for sizing — it's narrower (US-only, retail-only) and gives a tighter range that's actionable for capacity planning at the payment-rails layer.
Payment-rails authentication — Visa, Mastercard, AmEx, Cloudflare. Cloudflare announced a partnership with Visa, Mastercard, and American Express on 14 October 2025 on agent-authentication standards for transactions (Cloudflare press release, 14 Oct 2025 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-visa-trusted-agent.html). The technical foundation is Web Bot Auth — Cloudflare's verification approach using Ed25519 cryptographic signatures in HTTP Message Signatures (RFC 9421) headers, with a /.well-known/http-message-signatures-directory JWKS endpoint advertising public keys (Cloudflare developer docs — Web Bot Auth). Cloudflare's October 2025 follow-on blog post described how Visa's Trusted Agent Protocol and Mastercard's Agent Pay both build on Web Bot Auth, introducing a new tag in the Signature-Input header to indicate browsing vs purchasing intent and a nonce field for replay protection (Cloudflare blog, "Securing agentic commerce", 24 Oct 2025 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-secure-agentic-commerce.html). Visa's December 2025 newsroom release framed Trusted Agent Protocol as part of its broader Visa Intelligent Commerce initiative (Visa newsroom, 18 Dec 2025 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-visa-trusted-agent-protocol.html). The architectural meaning: the card networks are not letting agentic commerce route around them via Stripe + ACP. Instead they're moving up-stack — the agent identity layer sits above the payment rail and the networks intend to be the trust authority. For an investor: the agentic-commerce TAM gets split between the protocol layer (Stripe ACP, Visa Trusted Agent) and the clearing rail underneath — Stripe and OpenAI built the first plumbing, Visa pushed back hard at the identity layer in less than three months.
Research and lead-extraction — Perplexity Comet. Perplexity launched the Comet browser in July 2025 and the Comet Plus subscription in August 2025, with a $42.5M revenue-share pool and an 80/20 split favoring publishers, across three payout categories: direct traffic to publisher sites from the Comet browser, citations when publisher content appears in AI-generated answers, and agent-driven use when Comet's assistant relies on publisher content to complete user tasks (Subscription Insider, 28 Aug 2025 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-subscription-insider-comet-plus.html). Comet is the canonical example of the agent-as-client browsing-agent-pages-of-publishers workflow — and the Amazon-Perplexity dispute that erupted in late 2025 (Amazon's complaint that Comet was completing purchases on behalf of users while stripping Amazon's own discovery surface; PYMNTS, "Perplexity's Agents Crash Amazon's Gate") is the canonical disintermediation fight that everyone in retail-platform land is now watching.
Supply-chain integration — the enterprise A2A case. Less consumer-visible but probably the largest near-term TAM: enterprise procurement systems (SAP, Oracle, Workday, Coupa) exposing resident agents that other enterprises' procurement agents talk to. The A2A launch partner list — SAP, ServiceNow, Workday, Salesforce, Atlassian, Box, Cohere, Intuit, MongoDB — is overwhelmingly enterprise-SaaS, which is the signal: the immediate revenue line is B2B procurement, not B2C commerce. The "tens of millions of Graviton cores" Meta committed to AWS in April 2026 for agentic AI (covered in the agentic primer §3) is the corresponding infrastructure footprint; for every $1 of agentic commerce revenue, there's a multiple of $ in cloud spend on the agent runtime, and a fraction of that is the agent-to-agent traffic itself.
The single highest-stakes question the investor raised is what this means for web.com / Newfold (the call's reference to "Neufeld" was almost certainly Newfold Digital — note that web.com as a standalone brand no longer exists). Walking the player set in order:
Web.com (now Network Solutions) / Newfold Digital. Two structural facts are load-bearing. First, Newfold consolidated the Web.com brand into Network Solutions on 18 June 2025 — Web.com is no longer a brand, the customer base, the AI Website Builder, and the products all live inside Network Solutions (Newfold press release, 18 Jun 2025; CEO Christina Clohecy quote: "This strategic consolidation brings together decades of innovation, reliability, and top-tier support, offering customers a comprehensive experience under the globally recognized Network Solutions brand"). Second, Newfold's sister brand Bluehost shipped GatorClaw on 16 April 2026 — a no-code agent-builder platform on Bluehost VPS hosting, "built on the emerging OpenClaw ecosystem … purpose-built for always-on, agentic workloads" (Newfold press release, 16 Apr 2026). The strategic read: the agent-addressable web is a defensive opportunity for Newfold, not a pure threat. The SMB customer base (~7M across Bluehost + Network Solutions + HostGator + Crazy Domains + Yoast + the rest of the Newfold portfolio) already pays Newfold for the underlying hosting; adding "and your site has a resident agent" is an upsell on the existing relationship. The infrastructure tie to Oracle Cloud (the Bluehost OCI migration completes in 2026) gives Newfold capacity headroom for agent runtimes that exceed what shared-hosting traffic patterns would consume. The threat is the inverse: if the AI agent layer commoditises the site builder — if GoDaddy Airo or Wix's AI builder lets a small business stand up a fully agent-addressable site in five minutes — then the value of Newfold's domain + hosting + builder bundle erodes, and Newfold's defense is to be the one shipping that builder. GatorClaw is exactly that play.
GoDaddy. GoDaddy launched Airo.ai on 13 November 2025 — an agentic AI orchestrator that registers a domain, builds a Website Builder site, generates a logo and policies, and ships an online presence end-to-end through agent delegation. Six agents at launch (Airo Agent as orchestrator; Website Builder Agent; Domain Search & Registration Agent; plus three others), six more added on 25 November 2025 (GoDaddy press release, 13 Nov 2025). GoDaddy's stake is the more aggressive of the two — Airo wraps the entire SMB workflow under agentic AI rather than bolting an agent-builder onto existing hosting. The structural read is the same: domain + hosting + site builder is being absorbed into a single agent-orchestrated bundle; whoever lands first with the smoothest end-to-end onboarding has a one-shot land grab. GoDaddy is ahead by ~5 months. The question for the Newfold board is whether GatorClaw is enough — it's an agent-builder for sites that already exist, not a one-shot onboarding flow.
Wix and Squarespace. Both have shipped AI-first website-creation experiences (Wix AI Website Builder; Squarespace Beacon AI / Squarespace GPT) but neither has shipped a full agentic-orchestrator equivalent to Airo or GatorClaw as of May 2026. Squarespace's "AIO Scanner" — tracking mentions across AI platforms — is a defensive product for the publisher side (SEO for the agentic web), not an offering for the site builder. Wix is closer (more AI features in-bundle) but still pre-orchestrator. Both are exposed to the same compression: if SMB customers go straight from "I have a business idea" to "I have a deployed agent-addressable site" inside GoDaddy Airo or Newfold GatorClaw, the multi-step Wix/Squarespace onboarding becomes the legacy flow. The defensive move is to ship their own orchestrator; the offensive move is to position as the high-end design platform — that's Squarespace's natural ground but it shrinks the TAM at the bottom.
Shopify. Shopify is the most insulated of the SMB-platform group, for two reasons. First, Shopify shipped four official MCP servers (Storefront, Customer Account, Checkout in preview, Dev) in 2025–2026 and is a named NLWeb adopter — meaning every Shopify storefront is, by default, an agent-addressable surface. Second, Shopify is the platform OpenAI named as the canonical Instant-Checkout partner; the "over one million Shopify merchants" line in the Stripe/OpenAI launch post is the most economically valuable single commitment in the entire agentic-commerce protocol stack. The Amazon-vs-Perplexity dispute illustrates the asymmetry: where Amazon is fighting agent disintermediation, Shopify embraced it as a distribution channel. Shopify's underlying CMS, payments (Shop Pay), and fulfilment (Shopify Fulfillment Network) stay intact even when discovery moves to an agent; in fact, agent-routed orders increase Shopify's GMV without consuming Shopify's marketing budget. The structural position is comparable to AWS's in cloud — when the abstraction layer above moves (Kubernetes, then serverless, then agents), the underlying compute layer keeps growing.
Cloudflare and the CDN/edge layer. Cloudflare is the player making the most aggressive strategic bet of any incumbent infrastructure vendor. Three moves are stacked:
research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-pay-per-crawl.html). Publishers and partners that supported the launch include ADWEEK, The Atlantic, BuzzFeed, Fortune, O'Reilly Media, Quora, Stack Overflow, TIME, the News/Media Alliance, IAB Tech Lab, and Webflow, plus Gannett Media via a separate statement from its Chief Consumer and Product Officer Renn Turiano (Search Engine Land coverage, 1 Jul 2025 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-searchengineland-pay-per-crawl.html).research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-dynamic-workers.html), Sandboxes (GA), Cloudflare Mesh (private agent networking without VPNs), Agent Memory (managed durable agent state), AI Gateway as a unified inference layer across 14+ model providers (Cloudflare Agents Week recap · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-agents-week-2026.html).Cloudflare is positioning to be the agent infrastructure layer of the open web — the layer between the resident agents on web pages and the client agents reaching them. This is a credible play because Cloudflare already terminates a meaningful share of internet traffic at the edge and has the bot-fingerprinting graph to enforce authentication at scale. If Cloudflare succeeds, the company moves from CDN-and-security to agent-edge, with three new revenue lines (pay-per-crawl margin, Workers compute for agent runtimes, agent authentication as a metered service). If it doesn't, hyperscalers (AWS, Azure, GCP) absorb the agent-runtime layer into their existing cloud bills.
Akamai and Fastly. Akamai positioned itself differently — its 2025–2026 thesis is Edge AI Inference, announced October 2025 and reinforced in March 2026 with a Blackwell GPU buy across 4,400+ edge locations to run small language models close to the agent endpoint. Akamai is also disclosing 300%+ growth in AI bot traffic across its network, which is the corroborating volume signal (Akamai blog, "AI Pulse: How AI Bots and Agents Will Shape 2026"; both direct WebFetch and Wayback Machine return 403/no capture, metadata-cited). Fastly is a niche player here — its bot management is competent but it has not announced an agent-runtime or agent-authentication play distinct from Cloudflare's or Akamai's, which is the strategic exposure: Fastly risks being the edge-CDN that isn't an agent platform.
Hyperscalers — AWS, Azure, GCP. The hyperscalers' play is to absorb the agent-runtime layer entirely. AWS Bedrock AgentCore (announced June 2025, GA 2026) ships a managed agent-execution service with built-in MCP support; Azure AI Foundry shipped agent orchestration in mid-2025 and integrated NLWeb out of the box at Build 2025; GCP launched the Agent Builder in Vertex AI in 2025 with first-class A2A support (Google's own protocol). Cloudflare's bet is that the edge — not the hyperscaler region — is the right architectural place for the resident agent of a public-facing website (low-latency to end-user devices, low-latency to client agents that are themselves hyperscaler-hosted). Hyperscalers' counter-bet is that the resident agent should live next to the merchant's existing application stack, which is in the hyperscaler region. Both are plausibly right for different segments — high-volume public-web agents will likely run on Cloudflare-like edge; enterprise back-office agents will run in the hyperscaler region.
Search and discovery — Google, Bing, Perplexity, Brave. The disintermediation risk is real and quantified. Digiday's reporting on Google AI Overviews documents CTR collapse on category leaders — one lifestyle publisher's CTR fell from 5.1% to 0.6% over the past year, an automotive publisher from 2.75% to 1.71% — with the Digital Content Next consortium showing a median -10% YoY decline in referred traffic from Google Search overall (-7% news brands, -14% non-news), and UK PPA evidence showing 10–25% YoY CTR declines (Digiday, "Google AI Overviews" · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-digiday-ai-overviews.html). Search-referral traffic dropped 60% for small publishers (1K–10K daily page views) over two years; 47% for medium; 22% for large (ALM Corp / Chartbeat data). Google Search itself remains the volume layer, but the quality of remaining search traffic has compressed. ChatGPT's share of generative-AI web traffic has fallen materially through 2025–2026 as Gemini's share has risen — the discovery layer is consolidating around 2–3 winners. For an investor: ad-driven media companies are structurally exposed to the upstream cut; the open question is whether the agentic-commerce layer reroutes value back through identifiable agent-to-publisher payments (Comet Plus, Pay Per Crawl) or settles at a permanently lower equilibrium for content.
Cross-reference the agentic primer §5: the network section established that agentic clusters shift the east-west / north-south traffic mix toward east-west (server-to-server inside the data center) and introduce scale-across (NVIDIA Spectrum-XGS, August 2025) as a new fabric category for distributed AI super-factories. The agent-to-agent layer adds a second wave on top.
The east-west surge intensifies. A single agent task already touches multiple internal services (vector DB, tool servers, model endpoint) for one ReAct loop iteration. An agent-to-agent task at the same layer touches additional services: the calling agent's reasoning loop, the resident agent's reasoning loop, each side's tool calls (MCP), the discovery fetch, and the streaming session over SSE for status updates. The Cloudflare Agents Week 2026 framing of Cloudflare Mesh as a "private network layer for agents" is the structural admission that agent-to-agent traffic is becoming a distinct east-west category that needs its own fabric.
A new north-south signature. Where chatbot traffic was bursty north-south — short request, long response, long idle gap — agent-to-agent traffic is sustained north-south on long-lived SSE sessions with frequent JSON-RPC round-trips. Two seconds of SSE for a task progress update, then a 30-second pause, then another 5 seconds — repeated for the duration of a multi-minute task. This is closer to the WebSocket profile of the voice-agent case in the agentic primer §4.2 than to a traditional REST API call. The network buffering, idle-timeout, and connection-pooling assumptions in legacy load balancers were not designed for this pattern.
DC-to-DC pull-through accelerates. When the client agent lives in AWS us-east-1 and the resident agent lives in Cloudflare's London edge, every task traverses the public internet end-to-end. NVIDIA's Spectrum-XGS framing ("connect distributed data centers into giga-scale AI super-factories" — agentic primer §5) anticipated this for internal multi-DC AI training; the same architectural pressure now applies to external agent-to-agent traffic. The long-haul fiber capex thesis (TD Cowen on Ciena; JPMorgan on Dycom — agentic primer §5) gets a second tailwind from agent traffic distinct from the training-cluster east-west growth that drove the first.
Quantification is preliminary. Akamai's 300%+ AI bot traffic growth (Akamai blog, "AI Pulse"; WebFetch returns 403 and Wayback capture failed, metadata-cited) is the volume signal; Cloudflare's June 2024 measurement that AI bots accessed ~39% of the top 1M properties on its network while only 2.98% of those properties were blocking (Cloudflare blog, 3 Jul 2024 · local: research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-ai-bot-blocking-jul2024.html) is the addressable-market signal — most of the top 1M are reachable by AI bots today and most are not yet defended. No hyperscaler or analyst report yet sizes the agent-to-agent-specific network footprint as a distinct line item from the broader agentic-AI east-west surge — the BOM model's 06_AgenticAI archetype carries the 10.4% network + interconnect share, but the agent-to-agent fraction inside that is implicit. This is a near-term research gap worth filling.
Where new value accrues.
/.well-known/agent-card.json solves single-domain discovery; cross-domain agent discovery (an agent that needs to find any agent capable of "scheduling a service appointment in Chicago") needs a directory layer that doesn't yet exist as a paid service. The Linux Foundation's Agentic AI Foundation could end up here, or a startup could; Solo.io has flagged the gap publicly (Solo.io, "Agent Discovery, Naming, and Resolution"). Watch list: VC funding rounds for "agent registry," "agent DNS," "AgentSEO" categories in 2026.Where value is structurally compressed.
What's still TBD.
Six leading indicators for John's tracking through 2026–2027.
/.well-known/agent-card.json and the Schema.org-NLWeb headers Microsoft documented at launch.research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-google-a2a-launch.html…/2026-05-19-google-a2a-launch.htmlresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-a2a-agent-discovery-spec.htmlresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-microsoft-nlweb-launch.htmlresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-linuxfoundation-agntcy.htmlresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-linuxfoundation-aaif-formation.html — for the AAIF formation date, co-founders, and full platinum-member list…/2026-05-19-a2a-agent-discovery-spec.htmlresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-ai-bot-blocking-jul2024.html — for the 39% access rate / 2.98% blocking rate on the top 1M properties on Cloudflareresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-stripe-openai-instant-checkout.html — ACP mechanics, Will Gaybrick quoteresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-agenticcommerce-dev.html — for the Apache 2.0 license of ACP (verbatim: "ACP is open source and community-designed under the Apache 2.0 license")research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-visa-trusted-agent.htmlresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-secure-agentic-commerce.html — for Web Bot Auth + Ed25519 + HTTP Message Signatures + Visa Trusted Agent Protocol + Mastercard Agent Pay verbatim references/.well-known/http-message-signatures-directory JWKS mechanicsresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-visa-trusted-agent-protocol.html — for Visa Intelligent Commerce + Trusted Agent Protocol framingresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-subscription-insider-comet-plus.html — $42.5M pool, 80% publisher split, three payout categoriesresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-pay-per-crawl.htmlresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-content-independence-day.html — sibling Pay Per Crawl announcementresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-searchengineland-pay-per-crawl.html — for the named publisher/partner list (AdWeek, The Atlantic, BuzzFeed, Fortune, O'Reilly Media, Quora, Stack Overflow, TIME, News/Media Alliance, IAB Tech Lab, Webflow + Gannett Media via Renn Turiano quote)research/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-agents-week-2026.htmlresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-cloudflare-dynamic-workers.html — for the V8 isolate "100x faster than a typical container" verbatim claimresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-digiday-ai-overviews.html — for lifestyle publisher CTR 5.1% → 0.6%, automotive 2.75% → 1.71%, DCN median -10% YoY, UK PPA 10–25% YoY declines…/2026-05-19-cloudflare-agents-week-2026.htmlresearch/2026-05-19-agent-to-agent-primer/snapshots/2026-05-19-subscription-insider-comet-plus.htmlCross-references:
docs/briefs/2026-05-15-agentic-ai-primer.md (anchor for ReAct, tool-use schemas, MCP; cross-referenced as §2 throughout)docs/briefs/2026-05-15-token-primer.md (token economics that fund the protocol layer)models/dc-archetype-bom-2026-05-12.xlsx (sheet 06_AgenticAI for the network + interconnect cost share)