the web was built for eyes.
the next one is built for agents.
for thirty years we built the web for people looking at screens: pages to read, menus to click, forms to fill. now more and more of the looking is done by an agent, and an agent does not want a page. it wants to ask. this take is about what that changes: search, the shape of our data, software itself, and who stands behind every agent. some of it is being built right now, in pieces, by many people. some of it is still an open question, and those are the ones i want to work on in the open.
the visit to the website is fading. the question is not.
you already feel it. you ask claude or chatgpt, and the answer comes back with a few links you never open. pew research looked at 68,879 google searches and counted the clicks.
the page is not gone. agents still read it, and the index underneath search still does its job. what is fading is the trip. people stop going to the website, because their agent went for them.
and an agent reading a page made for a person is a strange sight. it wades through the menu, the banner and the cookie notice to find one sentence. it fetches the same page again and again, because it cannot tell what changed. over half of what ai crawlers fetch is pages they already had.
an agent does not want a page. it wants to ask.
search becomes a conversation between agents.
finding things on the web has already changed shape three times. each time, the thing in the middle got smarter, and you did less of the work.
- 1994 · a directorysomeone lists it by hand
two stanford students keep a list of sites they like, jerry and david's guide to the world wide web. it becomes yahoo.
- 1998 · an indexa machine reads everything
google crawls every page it can find, and orders them by the links pointing at them.
- 2024 · an answerit reads the pages for you
google's ai overviews summarise the pages, so you read the answer, not the page.
- next · a questionyour agent asks theirs
the answer comes live, from whoever knows it, not from a page that was right last week.
search today is three jobs, done by one company. it finds the candidates. it gives you an answer. and it decides the order. on the agent web, i think those three come apart. my reading
the middle row is the big one. a page can tell you a plumber exists. only the plumber can tell you he is free tomorrow at ten. live, personal questions, is she free thursday, what is your price today, can you deliver on friday, are better asked to the other side's agent than read from a page. here is how i picture a search like that.
- 01it looks in open lists
more than one, so you are never stuck with the one list that sold its top spot.
- 02it asks their agents, live
a few at a time, a short question each. nothing is shouted to everyone.
- 03it ranks for you
people you know and dealt with first, then what is hard to fake. never what a business says about itself.
- 04you decide
the agents draft and compare. the yes, the booking, the price, stays yours.
lists have a habit.
the first step does not go away. to know whom to ask, your agent still needs a list of who is out there. and lists have a habit. the yellow pages began as a plain list of every business in town. by 2007, businesses in the us were paying $14.3 billion a year in directory ads to stand out in it.
the open networks of the last few years felt the same pull. they were built to have no owner, and the list ended up owned by whoever paid for it. farcaster's biggest indexer bought the whole protocol in january 2026. when nostr's main search service went dark this autumn, apps found out by timeouts. so the lists of the agent web must be open, there must be more than one, and any one of them must be easy to replace.
one detail from nostr stays with me. when that shared search and spam list went dark, apps fell back on each person's own follows and mutes. the only trust that survived was the person's own. that is why the order belongs to your agent.
and seo? i think it loses its meaning. seo was the craft of helping a crawler understand a page that could not speak for itself. on the agent web, the things we used to crawl speak for themselves: the plumber's agent says when he is free, the shop's agent says what is in stock today. my reading
what comes after is a more efficient kind of search. i do not know its name yet. but it will have its own rules, the way seo once had, and people will learn to play by them. the question is who writes those rules. i want them written in the open, with your agent playing for you.
every agent has someone behind it.
in july 1993 the new yorker ran a cartoon by peter steiner. a dog sits at a computer and tells another dog: "on the internet, nobody knows you're a dog." for thirty years that was the joke, and the problem.
on the agent web it cannot stay a joke. there will be far more agents than people. one social network for agents counted 1.5 million agents, owned by about 17,000 people. that is about 88 agents for every owner, and no way to tell which ones a person stood behind.
so every agent has to answer one simple question: who do you act for?
about 88 agents for every owner: 1.5 million agents and about 17,000 owners, on one social network for agents.
i see two kinds of someone. a person, and an organisation: a shop, a clinic, a company. every agent hangs under one of them. and when an agent starts another agent to help, the helper is not a new citizen. it is one step further down the same chain, with less power than the step above. my reading
- youhold your keys, and give every yes.
- your agentdoes what you allowed, and says it is an agent.
- a helper it startsreads and drafts for it. never meets a stranger, never spends.
- the bakerywith a real person inside who answers for it.
- its agentanswers customers inside limits that person set.
- a helper it startschecks the stock, drafts a reply. a price or an order still waits for a person.
people and organisations should also be found differently. a bakery wants to be found. a person mostly does not want to be browsed. so on the agent web, i think you reach a person the way you reach a friend today: someone gives you their name, sends you an invite, or introduces you. an organisation can choose to be listed, with a real person inside who answers for what its agents say.
the law already leans this way. in 2024, air canada's chatbot told a grieving passenger that a bereavement fare could be claimed after the flight. the airline argued the chatbot was responsible for its own words. a tribunal disagreed, and air canada paid.
formats for the eye fall. shapes for agents rise.
in what collapses i wrote that the file format falls first. the pdf, the deck and the doc were made so everyone reads the same thing the same way. your agent can show you the work the way you understand it best. i still think that falls fast.
but when two people's agents share something, a task, a booking, a contact, they still have to agree on what it is. not how it looks. what it is: which part is the title, which is the due date, who may change the status.
today that agreement is buried in an api someone wrote years ago, between two apps that already know each other. on the agent web, i think it comes out into the open: a named shape both sides speak, that the model reads to understand and plain code uses to check. two pieces of software that never met can then share a list and keep it in step. my reading
the contract does not die. it moves, from code written once into a shape chosen each time two people share.
the closest thing i know is the barcode. on 26 june 1974, a cashier at marsh supermarket in troy, ohio scanned a ten pack of wrigley's juicy fruit gum. it was the first product sold by scanning a barcode. the barcode was not clever. it was agreed: makers printed it, shops read it, and both got something back, faster tills and stock they could count. a shared shape spreads when both sides gain from it.
the web taught the same lesson twice. schema.org, the markup that tells search engines what a page is about, reached 44% of domains, because google rewarded it with richer results. llms.txt, a file that tells ai what a site is about, sits on tens of thousands of sites, and 97% of them got no ai requests in one month this year, because nothing rewards it.
the parts for shapes between agents already exist, and that is good news. the at protocol behind bluesky finds named schemas through dns. ink & switch showed how one shape of a thing can be mapped onto another. automerge keeps two copies of a document in step with no server in the middle. what is missing is putting them in the hands of people's agents, with a reason for both sides to use them.
much of software becomes templates.
in 2009 apple sold the iphone on a promise: "there's an app for that." and there was. an app for the torch, the tip, the train times, the timer, the level on the shelf. most of them now fit in one sentence to your agent.
once the model is the intelligence, software does not need to be smart. no program is as smart as the model you already talk to. it needs to be scaffolding: plain files, clear tools and open rules your agent works through. here is where i think each kind of software goes. my reading
- an app for one job
your agent does the job itself, or rebuilds the app from what it does.
falls first - software many people share
it stays. your agent works it for you, and its screens matter less.
changes how it is used - the part between people
the api, the connector, the sync. it becomes a shape both agents speak.
becomes a template
the third kind is where the biggest change hides. most of what we pay software for is not the screen. it is the shared part: the board my team and i both see, the booking both sides trust. when that part is a template any agent can speak, shared work stops belonging to one app.
that is why openthings builds open software at all. opensoftware brings open tools to people through the ai they already use. scaffolding that speaks open shapes can work with anyone's, not only with its maker's.
the agent web is being built right now, in pieces.
this is not one company's project, and it should not be. look at what arrived in about two years, and where each piece sits between two people.
i am glad every one of these exists. it means the problem is real, and many people see it. most start from the agent, or tie the person in through one issuer, one email, or a server that can read the mail.
the piece i care about is the span in the middle: person to person, through your agents, with the yes on your own device. that is where OHAN, the open human and agent network, starts. and where someone has already built a piece well, OHAN should plug into it, not rebuild it.
what nobody has solved yet.
the hard parts of the agent web are not solved, by us or by anyone. these are the questions openthings is working on. i want to work on them in the open, with researchers, builders, and anyone who has watched a network go wrong.
- 01a yes others can check
when you approve something, can the other side verify that a person said yes, without learning who you are?
- 02software that never met
how do two agents share an item, keep it in step, and refuse what they should not accept, with no api between them?
- 03power that only shrinks
when an agent hands part of a job to another, how does every step carry less power, and some acts never pass on at all?
- 04finding without a centre
how does your agent find who can help among millions, when nobody owns the list?
- 05who stands behind an agent
how does an agent show that a person or an organisation is behind it, without one company becoming the gate?
- 06words that steer
how does an agent read a stranger's message without being steered by it?
- 07while you sleep
how does a message reach your agent when your laptop is shut, with no server that can read it?
- 08who keeps the shared parts
who runs the lists and relays everyone uses, who pays for them, and what keeps any one from becoming the next centre?
any answer has to keep a few rules. a person or an organisation stands behind every agent. the model drafts, and a person decides. no server reads your messages. nobody can take your name. revoking stops what comes next, and cannot take back what was already shared.
if one of these is your question too, let's connect.
it is not an app. it is a layer.
it is easy to look at agents talking to agents and see a new kind of messaging. it is bigger than that. the same few pieces, someone behind every agent, a name, a door you control and a yes only you give, carry everything else.
the web was not one company's product, and the agent web should not be either. if someone answers one of these questions better than i do, the open rules should take their answer. that is what open means to me.
agents are becoming the way we reach everything. who they answer to decides what kind of world that is. i want them to answer to people.
sources.
- clicks with and without an ai summary: pew research center, july 2025.
- over half of ai crawler traffic fetches unchanged pages: techcrunch, on cloudflare, july 2026.
- jerry and david's guide to the world wide web, 1994: wikipedia, yahoo. google and pagerank, 1998: wikipedia, pagerank. ai overviews, may 2024: google.
- us directory advertising, $14.3 billion in 2007: purple motes, from industry data. neynar buys farcaster: the block, january 2026. nostr's search service going dark, and apps falling back on follows and mutes: porwr #267, relay-outpost #343.
- "on the internet, nobody knows you're a dog", the new yorker, 5 july 1993: wikipedia. 1.5 million agents and about 17,000 owners: wiz, on moltbook. air canada and its chatbot: american bar association.
- the first barcode scan, 26 june 1974: wikipedia, universal product code. schema.org on 44% of domains: web data commons. llms.txt with no ai requests: ahrefs, may 2026; its spread: ppc land. named schemas through dns: at protocol lexicon resolver. mapping one shape onto another: ink & switch, cambria. automerge.
- "there's an app for that", 2009: wikipedia.
- the pieces: mcp; a2a; the NANDA draft at the ietf; aweb; x401, help net security; agentid, agentmail; the passkey to payment demo, tech coffee house.
- the readings marked "my reading", and the open questions, are mine.