AI Domains9 min read

Agentic web: what it is and what it means for your domain

What the agentic web is, which protocols run it, and why each of them still depends on a domain. We also report what the Forbes AI 50 publish for AI agents, from our check on October 22, 2026.

Terminal window running a curl request for a website's /.well-known/agent-card.json file, illustrating how AI agents discover sites on the agentic web

The agentic web is the part of the internet where AI agents, not only people, find services, read pages and complete tasks for a user. It runs on a few open protocols, such as MCP and A2A, and on plain files that sit on a domain. That last part is why your domain still matters.

Two dates set the scene. In August 2026, Google’s Agent2Agent protocol joined the Agentic AI Foundation, the Linux Foundation fund that has hosted Anthropic’s Model Context Protocol since December 2025. On October 7, 2026, ICANN published 13 applications for a new .agent extension. As of October 22, 2026, no .agent name can be registered.

What is the agentic web?

The agentic web is the phase of the internet in which AI agents act for people: they search, compare offers, fill in forms, call tools and hand work to other agents. The authors of the July 2025 paper Agentic Web: Weaving the Next Web with AI Agents describe it as “a new phase of the internet defined by autonomous, goal-driven interactions.” No standards body owns the term.

Microsoft put the phrase at the center of its Build 2025 conference. In a blog post on May 19, 2025, the company’s Frank X. Shaw described “an open agentic web, where AI agents make decisions and perform tasks on behalf of users or organizations.” Read it as one company’s vision rather than a specification.

Agentic means able to get results with autonomy, according to Merriam-Webster, and the sense is now used mostly for AI. Plain ChatGPT chat, which answers questions, is not agentic because it does not act on its own. OpenAI’s ChatGPT agent, launched on July 17, 2025, is agentic: according to OpenAI’s release notes, it runs its own virtual browser, opens sites and fills in forms.

AI search and the agentic web overlap, but they are not the same thing. In AI search, a model reads your page and writes an answer, which is the subject of our review of generative engine optimization research. In the agentic web, the model also takes the next step: it books, buys, files a form or asks another agent to do it.

The protocols behind the agentic web

Three open protocols carry most of the agentic web as of October 22, 2026. Anthropic’s Model Context Protocol (MCP, November 2024) connects an AI model to tools and data. Google’s Agent2Agent protocol (A2A, April 2025) lets agents find and talk to each other. Microsoft’s NLWeb (May 2025) gives a website a natural-language endpoint that is also an MCP server.

Anthropic introduced MCP on November 25, 2024 as an open standard for “secure, two-way connections” between data sources and AI tools. Google announced A2A on April 9, 2025, with more than 50 partners at launch, and gave it to the Linux Foundation on June 23, 2025. NLWeb came from R.V. Guha, the creator of RSS, RDF and Schema.org, and Microsoft’s NLWeb announcement says the company believes NLWeb can play “a similar role to HTML in the emerging agentic web.”

MCP and A2A now share one neutral home. The Linux Foundation formed the Agentic AI Foundation on December 9, 2025, with MCP as a founding project next to Block’s goose and OpenAI’s AGENTS.md. A2A joined the same foundation in August 2026, and its project blog says more than 150 organizations back it.

Two lighter pieces sit beside the big three. Jeremy Howard proposed llms.txt on September 3, 2024: a Markdown file at the root of a site that points language models to the pages worth reading. It is a proposal, not a standard. WebMCP, which Chrome previewed on February 10, 2026, lets a web page offer tools to a browser agent from inside the page.

ProtocolMade by, yearWhat it connectsWhere it lives on a domainStatus, October 22, 2026
MCPAnthropic, 2024Agent to tools and dataA server URL; optional /.well-known/ai-catalog.jsonOpen spec, hosted by the Agentic AI Foundation
A2AGoogle, 2025Agent to agent/.well-known/agent-card.jsonSpec version 1.0.0, hosted by the Agentic AI Foundation
NLWebMicrosoft, 2025Website to agent, in plain languageAn endpoint on the site that is also an MCP serverOpen-source project
WebMCPGoogle and Microsoft, 2025Web page to browser agentA JavaScript API inside the pageOrigin trial in Chrome 149
llms.txtJeremy Howard, 2024Site content to language models/llms.txtCommunity proposal
Comparison of four agentic web building blocks, MCP, A2A, NLWeb and llms.txt, by maker, year, purpose and where agents discover them on a domain
MCP and A2A connect agents to tools and to each other, while llms.txt is the simplest file a site can publish. The MCP discovery file is optional. Source: Anthropic, Google, Microsoft and llmstxt.org documentation, as of October 22, 2026.

Why domains still matter for AI agents

Agent discovery still runs on domain names. An A2A agent publishes its Agent Card at https://{domain}/.well-known/agent-card.json, a path IANA has listed as a permanent registration since August 1, 2025. The official MCP Registry can tie a server’s name to the publisher’s domain and verify ownership through a DNS TXT record or a /.well-known file on that same domain.

The A2A specification names the agent card path directly and says cards should be served over HTTPS. A card can also carry a JWS signature, so a client can check that the card was not altered and comes from the claimed provider. Older A2A versions used agent.json; the current version 1.0.0 uses agent-card.json.

The MCP Registry makes the link even plainer. Under its domain authentication rules, a company at example.com publishes servers under the reverse-DNS name com.example. To prove it, the company adds a TXT record on its apex domain or a file at /.well-known/mcp-registry-auth, each holding the company’s public key. The registry trusts whoever controls the domain.

MCP’s core specification has no discovery file. An optional Server Card extension, finalized in October 2026, lets a domain list its MCP servers in an AI Catalog at /.well-known/ai-catalog.json. As of October 22, 2026, though, agents usually reach an MCP server through its URL or a registry. IETF individual drafts go further and propose agent discovery in DNS itself, for example through records under an _agents label. They are drafts, not standards.

The practical result: each of these files and records sits on a domain, so an agent’s identity is only as stable as the domain behind it. If the domain lapses, whoever registers it next controls the paths that agents read.

What the Forbes AI 50 publish for agents

On October 22, 2026, BuySI Domains checked the main websites of all 50 companies on the Forbes AI 50 2026 list for agent-facing files. Of those 50, 30 served a valid llms.txt file, and seven of them also served llms-full.txt. Three served some MCP discovery file, one of them at /.well-known/ai-catalog.json. None served an A2A agent card at /.well-known/agent-card.json. On each of these checks, a bot challenge blocked two sites, three different sites in all.

The method was simple and can be repeated. A script requested seven paths on each main domain in two passes 40 seconds apart, and both passes gave the same counts. A file counted only if it returned real content: a JSON agent card with a name and skills or capabilities, or an llms.txt that opens with a Markdown heading. A path that sent back the homepage counted as missing.

We checked main domains only, not docs or developer subdomains, so the counts are a floor. Paths behind a Cloudflare challenge count as unknown, not as missing.

Bar chart of agent-facing files on 50 Forbes AI 50 2026 main websites: llms.txt 30, llms-full.txt 7, MCP discovery file 3, A2A agent card 0
Most Forbes AI 50 companies publish llms.txt, but none serves an A2A agent card on its main domain. Source: BuySI Domains check of 50 Forbes AI 50 2026 websites, October 22, 2026.

Effort explains much of the spread. An llms.txt file is a single text file that needs no new software, and most of the list publishes one. An agent card only makes sense for a company that runs an A2A agent for the public, and none of these 50 publish one at the root of their main domain. This check shows who publishes llms.txt, not which AI systems read it. Our guide to LLM optimization covers what is known about how brands get into AI answers.

.agent is years away; .si works today

ICANN’s 2026 new gTLD round drew 13 applications for .agent, according to the data ICANN published with its October 7, 2026 reveal. Applicants include OpenAI, Meta and Charleston Road Registry, Google’s registry company. ICANN has set no launch date for .agent, and nothing under .agent can be registered as of October 22, 2026.

The .agent extension is years from launch, and ICANN’s own schedule shows why. According to the ICANN reveal announcement of October 7, 2026, String Confirmation Day is November 17, 2026. The comment and objection window runs to March 16, 2027, and string evaluation is expected to take about 180 days. With 13 applicants for one string, .agent must also go through contention resolution, and only one applicant can run it.

Identity Digital, a registry operator, estimated in February 2026 that the first uncontested new extensions could launch in late 2027 or early 2028. Contested strings fall outside that estimate. Our overview of the 2026 new gTLD round covers the process and the other AI strings in it. Names sold as .agent today on alternative roots are not ICANN domains.

.si is Slovenia’s country-code domain. For AI companies it reads as SuperIntelligence. That reading is our framing, not an official meaning. Unlike .agent, .si is live. Clause 2.7 of the Register.si general terms defines a holder as “a natural or legal person who registers a domain name” under those terms, with no residency condition stated.

Every file in this article, whether an agent card, an llms.txt file or an MCP Registry TXT record, works on any domain you control and needs only DNS and HTTPS. A short .si name such as essai.si, which carries both “AI” and “SI”, can host them the day it resolves. BuySI Domains lists essai.si for sale. For the rules and setup, see the complete .si domain guide, and for the naming side, whether .si suits an AI startup.

What to do on your domain now

For most sites, the agentic web calls for a few small, reversible steps on a domain you already own. The order below runs from cheapest to most involved.

  1. Publish llms.txt if you have docs or product pages. It is one Markdown file at /llms.txt. Treat it as a low-cost proposal, not a ranking factor.
  2. Add an agent card only if you run an A2A agent. Serve it at /.well-known/agent-card.json over HTTPS, and sign it if your tooling supports JWS.
  3. Claim your MCP namespace if you ship an MCP server. The MCP Registry checks a DNS TXT record or a /.well-known/mcp-registry-auth file on your domain.
  4. Treat the domain as infrastructure. Turn on auto-renew, lock transfers at the registrar and give DNS access to at least two people.
  5. Watch .agent, do not wait for it. The next ICANN milestone is String Confirmation Day on November 17, 2026. Skip .agent offers that sit outside ICANN.

A startup naming an agent product today can launch on any live extension, publish its files there and decide on .agent if and when it opens. A company that already ships an MCP server or an A2A agent has the most to gain from steps 2 and 3, because both tie its agent to the domain it already owns.

Frequently asked questions

What is the agentic web?

The agentic web is the part of the internet where AI agents act for people: they find services, read pages, fill in forms, call tools and pass tasks to other agents. A July 2025 arXiv paper describes the agentic web as a new phase of the internet defined by autonomous, goal-driven interactions. It runs on open protocols such as MCP and A2A and on files served from ordinary domains.

Is ChatGPT an agentic AI?

Agentic means able to get results with autonomy. By that test, ChatGPT used as a plain chat that answers questions is not agentic, because it does not act on its own. OpenAI's ChatGPT agent, launched on July 17, 2025, is agentic: according to OpenAI's release notes, it runs its own virtual browser, opens websites and fills in forms to complete a task for the user.

What does agentic mean?

Agentic means able to get results with autonomy, according to Merriam-Webster. In AI, an agentic system pursues a goal on its own: it chooses steps, uses tools such as a browser or an API, and acts for the user instead of only answering a prompt. OpenAI's ChatGPT agent, launched on July 17, 2025, is one example. The agentic web is the part of the internet where such agents operate.

What is the difference between MCP and A2A?

MCP and A2A solve different problems. Anthropic's Model Context Protocol (MCP), released in November 2024, connects an AI model to tools and data sources. Google's Agent2Agent protocol (A2A), announced in April 2025, lets separate agents find each other and work together. As of October 22, 2026, MCP and A2A are both hosted by the Linux Foundation's Agentic AI Foundation.

Can I register a .agent domain?

No .agent domain can be registered as of October 22, 2026. ICANN's 2026 new gTLD round drew 13 applications for .agent, from applicants including OpenAI, Meta and Google's registry company, and only one of them can run the extension. ICANN has set no launch date. Names sold as .agent on alternative roots are not ICANN domains. An agent card or llms.txt file works on any existing domain today.

What is an agent card?

An agent card is a JSON file that describes an A2A agent: its name, skills and how to reach it. The A2A specification places the agent card at https://{domain}/.well-known/agent-card.json, a path IANA registered as permanent on August 1, 2025. Agent cards should be served over HTTPS and can carry a JWS signature. Older A2A versions used agent.json instead.

Sources

  1. ICANN reveals 2026 round applications for new generic top-level domains — ICANN
  2. Agent2Agent (A2A) Protocol Specification — A2A Project, Linux Foundation
  3. MCP Registry: authentication — Model Context Protocol (GitHub)
  4. Agentic Web: Weaving the Next Web with AI Agents — arXiv
  5. Introducing the Model Context Protocol — Anthropic
  6. Linux Foundation announces the formation of the Agentic AI Foundation — Linux Foundation
  7. A new chapter for A2A: joining the Agentic AI Foundation — A2A Project
  8. Introducing NLWeb: bringing conversational interfaces directly to the web — Microsoft
  9. Microsoft Build 2025: The age of AI agents and building the open agentic web — Microsoft
  10. The /llms.txt file — llmstxt.org
  11. General terms and conditions — Register.si
  12. Your roadmap to the next round: key dates and what to expect — Identity Digital