Can AI Agents Give Us the Internet Back?

Posted on October 20, 2026 by Evgeny Poberezkin
Tags: ,

They can – but not on the networks most people use.

Internet: a market that nobody owned

When the internet began, only research and education were allowed on it. In 1992 Congress opened it to commerce, and in 1993 CERN gave up all rights to the web’s code1.

Anybody could build on it, publish and read, and people found services in catalogs2.

How we lost it to big tech

Catalogs were curated manually and could not keep up as the web grew. Search replaced them, and a few companies came to own it. And software was expensive to build at scale, so well-funded companies built the services most people used.

The market became a few walled gardens. Their owners decide who gets found and charge for it3, see every buyer and take a cut of the deals.

AI changes the game

AI can solve the three problems that closed the internet market:

  • catalogs can scale again, because AI-powered services can curate and search them;
  • search indices can become a public good, because data is now cheap;
  • creating software costs a fraction of what it did, so anybody can build a service.

Agents can also make buying cheaper, safer and more convenient: an agent can compare a thousand times more offers than a person, without an aggregator skewing the comparison toward whoever pays; it can protect its owner from manipulation; and it gives everybody what only a few could afford – a personal concierge.

But AI also breaks trust in the networks we use. It can clone a voice and a face, and phone numbers and email addresses are easy to spoof. Anybody who has your number or email can target you, or pretend to be somebody you know – and an agent can be deceived the same way.

What keeps a market open

A market stays open only while three conditions hold:

  1. Sellers can be found and trusted. When anybody can be faked, a seller’s reputation should be attached to a trusted name, or agents will buy from impostors.
  2. Buyers can ask without being identified. A request tells who wants what, when and for how much. Whoever learns that can raise the price for the buyer who needs it most4.
  3. Nobody sits in between. A middleman decides who gets found and takes a cut of every deal – this is how the open market closed last time.

Why today’s solutions fail

Platforms. We saw what happens when a platform owns the market: it decides who gets found, sees every buyer and takes its cut. Agents cannot fix it. Meta’s business agents, for example, already sell inside WhatsApp and Messenger5, where Meta knows who buys what.

Websites with MCP services. An agent can call a seller’s website directly, but every request shows its IP address, and to get a reply later it needs an account – a registration that is hard for an agent, and that identifies the buyer. And it still finds sellers through search that big tech owns.

A public square. Rosenblat, Strauss and O’Reilly propose a public space where agents broadcast their requests, and sellers’ agents answer6. But broadcasting makes every request public, so buyers can no longer ask without being identified. The authors list privacy loss among the risks themselves. It should be the opposite: offers are public, and requests are private.

We built SimpleX for people – it is ready for agents

We have been building SimpleX since 2020 for private communication between people. It is the only messaging network without any user identifiers7: nobody can contact you, or pretend to be somebody you know, unless you give them a link. Over 600,000 people use it every month.

SimpleX can enable an open market for agents. Sellers would publish their addresses under SimpleX names that only they control8. Buyers’ agents would read sellers’ channels privately9 and contact each seller from a throwaway profile with a hidden IP address. And nobody sits in between, including us10: the servers relay encrypted messages without knowing who sends them to whom.

Some businesses already support their customers via SimpleX. Test names work now, and .simplex names go on sale on December 12, 2026. We have just released an MCP server11 that connects any agent to SimpleX, and published open specifications for websites and agents12.

What the market can become

On this network, the market can open again: a buyer’s agent compares thousands of offers, asks the best sellers privately, and buys with no platform taking a cut. To get there, we are building:

  • catalogs of sellers on SimpleX names, curated by people and AI-powered services;
  • agents calling each other’s tools with a request over SimpleX, without creating a conversation;
  • custom interfaces inside chats; and
  • an autonomous namespace, so that no company, including ours, can take a seller’s name away.

Join us and own it

Whatever you create and build on SimpleX – a community, a channel, a service or an agent – is yours. Tell us what you need from the network, and what is missing: write to us on SimpleX.

To own a part of the company that builds it, invest in SimpleX Chat’s community round on Wefunder.

  1. The NSFNET backbone policy prohibited its use “for purposes ‘not in support of Research and Education’” (A Brief History of the Internet). Public Law 102-476 of October 23, 1992 allowed networks to be used “substantially for purposes in addition to research and education” (section 4). On April 30, 1993, CERN declared: “CERN relinquishes all intellectual property rights to this code” (Licensing the Web).↩︎

  2. The best known was The Whole Internet User’s Guide & Catalog, published by Tim O’Reilly’s company in 1992; it sold more than a million copies (Brian Erwin, O’Reilly, 2016). In 1993 the company launched GNN – in O’Reilly’s words, “the first commercial website” (opening remarks at Bellagio, April 28, 2026).↩︎

  3. Paid placements fill the top results, and people click them even when the products are more expensive or worse rated: Rock, Strauss, O’Reilly and Mazzucato, Behind the Clicks: Can Amazon allocate user attention as it pleases?, UCL IIPP working paper 2023-11.↩︎

  4. The US Federal Trade Commission found in an initial analysis that details such as a person’s precise location or browser history are used to charge different people different prices for the same goods and services: FTC Surveillance Pricing Study Indicates Wide Range of Personal Data Used to Set Individualized Consumer Prices, January 17, 2025.↩︎

  5. “more than 1 million businesses using them to talk to their customers or complete sales every week”: Meta, Q2 2026 prepared remarks, July 29, 2026.↩︎

  6. Sruly Rosenblat, Ilan Strauss and Tim O’Reilly, Agents Need a Public Square, June 5, 2026.↩︎

  7. No phone numbers, usernames, account IDs or even random IDs. Instead of identifying users, SimpleX assigns random identifiers to each connection – separate ones for sending and receiving – so only the users know who they talk to: SimpleX whitepaper.↩︎

  8. SimpleX public names: “Only you control the name with the key in your wallet.”↩︎

  9. SimpleX channels: “Anybody can join a public channel via its link and see what is sent, but not who sent it, and not who else is reading.”↩︎

  10. SimpleX Chat licensed the network protocols to the SimpleX Network Consortium – an agreement between the nonprofit SimpleX Network Foundation and SimpleX Chat for protocol governance.↩︎

  11. Start an agent with npx -y @simplex-chat/mcp serve my-agent and join its chat in the SimpleX app: the agent asks there before it connects to anybody or sends a message, and you approve with one tap. It works with Claude Code now: @simplex-chat/mcp.↩︎

  12. A website marks its SimpleX addresses with the simplex link relation and lists them in /.well-known/simplex.json (Publishing SimpleX addresses on websites). Agents call each other’s tools with ordinary MCP messages, sent in a chat with /_call or as a single request (Agent protocols over SimpleX).↩︎