The Agentic Web Is Splitting Into Two Bets: Identity And Capability

The protocol layer of the agentic web is splitting into two bets, and most web sites have already positioned one in all them with out realizing it. The primary guess is about id: a file known as llms.txt that tells AI fashions who you might be and what your content material covers. The second is about functionality: a browser normal known as WebMCP that tells an agent what it will probably truly do in your web site as soon as it arrives. They sound like the identical concept. They reply reverse questions. And the distinction decides the place your effort ought to go over the subsequent yr, as a result of one in all these bets is being switched on for you by default, whereas the opposite, the one that truly lets an agent end a job, takes deliberate work.

2 Information, 2 Questions

An agent that lands in your web site has two issues it’d wish to know. The primary is what this place is and what it covers. The second is, now that it’s right here, the way it completes the duty its person despatched it to do. Id, then functionality. The agentic internet is constructing a separate reply for every, and the 2 solutions come from two totally different information backed by two totally different camps.

The id reply is llms.txt, a plain-text file you publish on the root of your area. It’s a curated map: right here is who I’m, listed here are my most vital pages, here’s what each is about. The potential reply is WebMCP, a browser normal that lets your web site expose callable instruments to an agent, so the agent can search, filter, value, or guide by invoking a perform as a substitute of guessing at your interface.

Llms.txt and WebMCP get lumped collectively as two flavors of the identical factor, agent-readiness. They reply two totally different questions. Id is a brochure. Functionality is the money register. And proper now the trade is pushing each web site towards the brochure, the guess with the weaker proof behind it, whereas the money register, the guess that truly strikes a transaction ahead, is left to the few groups keen to construct it.

The Id Guess: LLMs.txt

Llms.txt is a markdown file that lives at yourdomain.com/llms.txt. It was proposed by Jeremy Howard, co-founder of Answer.AI, on September 3, 2024, as a strategy to hand language fashions a clear, human-curated index of your content material as a substitute of constructing them crawl and reconstruct it from HTML cluttered with navigation, advertisements, and scripts. The pitch is cheap on its face. Fashions work higher with construction, so give them construction.

The issue is that the proof it does something is skinny. Answering folks on Reddit in June 2026, Google’s John Mueller known as llms.txt “purely speculative for now (the file has existed for years, but not one of the AI techniques use it)” (reported by Search Engine Journal). That may be a Google Search advocate saying the quiet half out loud. A file that has been round for over a yr. A file no main AI system has confirmed it reads. A file that asks you to maintain a second copy of your individual content material in sync.

Even with proof this skinny, llms.txt is being switched on for web sites by default. AIOSEO, a plugin used on greater than 3 million WordPress web sites, generates an llms.txt by default. A really massive variety of web site house owners are actually publishing an llms.txt they never decided to publish, describing their web site in a file they’ve by no means learn.

To be clear, I’m not in opposition to the file. A well-maintained llms.txt is a effective content material index. The issue is the default. The id guess is successful on adoption, not as a result of it really works, however as a result of a plugin checkbox shipped set to on. A file you didn’t write, that no AI system is confirmed to learn, that drifts out of sync with the web site it claims to explain the second you overlook it exists, is a wierd factor to have on the heart of an agent technique. However that’s the place quite a lot of web sites now sit, by default fairly than by alternative.

The Functionality Guess: WebMCP

WebMCP, quick for Net Mannequin Context Protocol, is a browser normal that lets your web site register callable tools an agent can invoke by means of a navigator.modelContext API. It begins from a special query than the id file: not who you might be, however, on condition that an agent is already in your web site, the way it correctly completes the duty it got here to do. That’s near how Mueller framed his personal choice in the identical dialog. He mentioned he likes the WebMCP method and the commerce integrations as a result of “they’ve clear targets & processes.”

As a substitute of an agent screenshotting your page and guessing where to click, your web site tells it instantly: listed here are the actions I assist, listed here are the inputs each wants, here’s what you get again. The management inverts. The web site declares its capabilities fairly than forcing the agent to reverse-engineer them.

The usual is being written by means of the W3C Net Machine Studying Group Group by engineers from Google and Microsoft, and it was printed as a draft on February 10, 2026. It’s now in a public origin trial that runs from Chrome 149 through Chrome 156. An origin trial is how Chrome lets an internet site activate an experimental function for its actual guests for a restricted window, by registering for a token, as a substitute of hiding it behind a developer-only flag. So you’ll be able to run WebMCP instruments on reside visitors now, not simply by yourself machine. The agent consuming these instruments in Chrome at the moment is Gemini.

WebMCP is the stronger guess as a result of it has a transparent job. An id file describes you and hopes somebody reads it. A functionality device will get invoked, returns structured information, and strikes a job one step nearer to achieved. The brochure waits to be learn. The money register rings.

Why The Two Bets Are Not Interchangeable

The cleanest strategy to maintain the distinction is thru the 2 design constraints I hold coming again to in Machine-First Architecture: id and interplay. Id is making your model and your content material unambiguously machine-readable. Interplay is letting an agent full an motion with a predictable outcome. Llms.txt is a play on the id layer. WebMCP is a play on the interplay layer. They’re instruments for 2 totally different jobs, and an internet site can want each, one, or neither, relying on what it needs brokers to do.

This issues extra now that agents are not a rounding error in your traffic. Cloudflare’s Chief Executive Matthew Prince said in June 2026 that automated visitors had handed human visitors for the primary time, at 57.3% of requests to webpages in opposition to 42.7% from folks, a crossover he had projected at SXSW wouldn’t arrive till 2027. When many of the requests hitting your web site are machines appearing for folks, whether or not the agent knew who you might be stops being the attention-grabbing query. Whether or not the agent might truly do what your buyer requested turns into the entire recreation. Id with out functionality is a machine studying your signal after which standing at a door it can not open.

That’s the inversion the default-on adoption will get backward. The file being switched on for everybody solutions the query that issues least as soon as the agent has arrived. The usual that solutions the query that issues most is the one nearly no person has turned on.

I Positioned Each Bets On My Personal Web site, Intentionally

I run both on No Hacks, and the distinction between how a lot each prices to position is the entire argument in miniature.

For id, I hold an llms.txt, and I generate it from the web site’s personal content material every time I replace the positioning, fairly than leaving a plugin to write down one I by no means learn. That’s the deliberate half. As a result of I rebuild it every time the positioning adjustments, it stays in keeping with what is definitely right here, and if an AI system ever does begin studying llms.txt, mine might be prepared when that day comes. I’m not betting on it. I’m conserving it present and low-cost and never mistaking it for a method.

For functionality, I carried out WebMCP. When a WebMCP-capable browser masses the web site, it registers 4 callable instruments by means of navigator.modelContext. Two cowl the glossary: one lists each time period so an agent can uncover what’s outlined, and one returns the canonical definition of a time period with its supply hyperlink. Two cowl my working panorama of agentic browsers and agent merchandise: one enumerates each tracked product by class, and one returns the total element on a single product by identify. An agent doesn’t need to scrape my pages to reply what I imply by the aptitude guess or which agentic browsers I monitor. It calls a device and will get a clear, structured reply.

The element that issues most is the place the instruments learn from. Each pulls from the identical information that powers the human-facing web page. The glossary instruments learn the glossary. The product instruments learn the product listing. So the agent reply and the human reply can by no means disagree, as a result of one supply feeds each. That’s the distinction between exposing a functionality and sustaining a separate description of your self, and it’s why the aptitude guess, constructed this fashion, has no separate copy to maintain in sync.

So I maintain each information. One took a day and does actual work. The opposite I regenerate from my very own content material and deal with as a hedge, not a plan.

What This Means For Your Web site

First, discover out what you will have already printed. Open yourdomain.com/llms.txt in a browser. If one thing masses, your stack put it there, presumably a plugin default you by no means set. Learn it. Ask whether or not it actually describes your website, as a result of if an AI system ever does begin studying llms.txt, an inaccurate one is worse than none in any respect. This prices 5 minutes, and most web site house owners have by no means achieved it. If you wish to hold one, generate it from your individual content material so it can not drift.

Second, determine whether or not the aptitude layer is value putting for you. If brokers have any cause to finish a job in your web site, search stock, check a price, start a booking, begin a return, then WebMCP is the guess that pays, and the origin trial means you’ll be able to run it on actual guests now fairly than ready for it to ship. In case your web site is solely one thing brokers learn fairly than act on, the aptitude layer can wait, and your effort belongs in clear, server-rendered content material that any agent can extract.

What you shouldn’t do is assume the file your plugin switched on is your agent technique. It’s an id declare no AI system is confirmed to learn, and it says nothing about what an agent can do as soon as it arrives.

What Is Nonetheless Unsettled

The trustworthy uncertainty: The id guess isn’t lifeless; it’s unproven. If a significant AI system declares tomorrow that it reads llms.txt and weights it, the calculus shifts, and the file everybody defaulted into all of a sudden earns its place. I’m awaiting precisely that sign and haven’t seen it but.

WebMCP has its personal open questions. It’s a Group Group draft, not a ratified normal. Gemini in Chrome is the principle agent consuming it up to now. And cross-browser assist is thinner than the secondary protection suggests. Microsoft co-authored the usual, however I couldn’t discover WebMCP in Microsoft Edge’s official launch notes as of model 147, so I might deal with any declare that Edge ships it natively as unconfirmed for now.

Till that sign arrives, my guess is evident: by the tip of 2026, functionality is the guess that issues, and id fades to a hedge. The file that hopes to be learn loses to the device that will get invoked, as a result of brokers are measured on whether or not they completed the duty, not on whether or not they learn your signal first. Place the aptitude guess intentionally. Generate the id file from your individual content material, when you hold it in any respect. And browse the one your plugin already wrote earlier than it speaks for you.

Extra Sources:

This put up was initially printed on No Hacks.

Featured Picture: Master1305/Shutterstock