The Agent Internet

Published March 17, 2026, by Frans

The web's biggest new user base since mobile isn't human.

The internet was built for people sitting at screens. Every protocol, every interface, every business model assumes a human is on the other end, someone with eyes to read, hands to click, and attention to sell. Search engines index pages for human readers. Websites render HTML for human browsers. Advertising prices human attention by the thousand impressions.

That assumption is breaking.

AI agents, software that acts on behalf of humans across the internet, are becoming a primary participant on the internet. Not for every interaction, not all at once. But the trajectory is unmistakable: the things people do on the internet (search, compare, purchase, schedule, communicate, manage) are being delegated to agents that do them faster, cheaper, or simply to save humans from having to do them at all.

This paper lays out the foundational framework for understanding what changes when agents become first-class participants on the internet.

The Shift: From Browse to Delegate

The human web is built around browsing: a human navigates to a page, reads it, makes a decision, takes an action. The entire UX stack exists to facilitate this: navigation menus, search bars, product grids, checkout flows, confirmation emails.

The agent internet is built around delegation: a human states an intent (e.g., "find me the cheapest flight to Tokyo next month"), and an agent fulfills it across whatever services are necessary, without the human navigating anything.

This is not a minor UX change. It restructures every layer of the stack, which I propose with A.I.R. to solve by:

Every row in that table represents an industry being restructured.

Three Phases of the Transition

Phase 1: Agents Use the Human Web

Agents interact with websites built for humans. They render pages, read DOM, click buttons, fill forms. This works, but it's fragile: a CSS change can break an agent's workflow. It's the equivalent of a human using a website through a screen reader designed for a different language. Functional, but inefficient.

This is where agent-web interaction began. Screen-scraping, DOM automation, headless browsers. It works well enough to be useful, poorly enough to leave enormous room for improvement.

Phase 2: The Dual-Stack Web

Websites begin serving two audiences simultaneously: humans (via HTML) and agents (via structured capability declarations, I propose an open standard called agent.json). A website publishes both its human-facing pages and a machine-readable manifest of what agents can do there.

This is analogous to how websites serve both HTML and RSS, or both web pages and API endpoints. The human interface does not go away. It gains a parallel structured layer that agents can consume directly.

The key innovation in this phase is the capability manifest: a standardized declaration of what actions are available, what inputs they require, and what outputs they produce. Instead of an agent reverse-engineering a checkout flow by reading HTML, the service provider declares: "Here is how to purchase a product. Send me a product ID and a shipping address."

Phase 3: Agent-Native Services

Some services will be built agent-first. No human UI at all, just structured capability endpoints designed for agent consumption. These services will exist because agents create demand for capabilities that no human would navigate to directly.

Example: A price comparison service that exists only as an agent capability. No website, no app, no UI. Agents query it as part of a purchase workflow. It monetizes through completion fees, not advertising.

This phase follows naturally from the first two, and the economic incentives are clear: if most of your "users" are agents, why build a human UI at all?

The Runtime Layer

Between the agent and the internet, there must be a runtime: an execution environment that translates agent intentions into real web interactions, manages browser state, handles authentication, enforces human oversight, and routes requests to the best available service.

This runtime layer has no precedent in the human web. It does not exist in the human web because the human was the runtime. A person's brain does the intent resolution, our eyes do the content extraction, our hands do the form filling, our judgment decides which service to use.

The agent internet requires this to be software. And that software must solve several hard problems simultaneously:

1. Universal compatibility. The runtime must work with every website, not just ones that have opted into agent support. The web has billions of pages. The runtime cannot wait for them all to publish capability manifests. This is how I designed A.I.R.

2. Progressive enhancement. When a website does publish structured capabilities (such as agent.json), the runtime should use them, because they're faster, more reliable, and cheaper. But the absence of structured support cannot be a blocker.

3. Human oversight. A person delegates a task, not their judgment. The runtime must enforce supervision boundaries: read actions proceed freely, write actions require approval, sensitive actions require explicit consent (dependent on user preferences and runtime policies).

4. Multi-agent coordination. A single task may require multiple agents working together, one researching, one comparing, one executing. The runtime must isolate their state while enabling coordination.

5. Economic settlement. When an agent completes an action on a service provider's behalf, value was created. The runtime must have mechanisms, internal or external, to track, attribute, and settle that value.

Who Benefits

Human Users

The obvious beneficiary. Instead of spending 45 minutes comparison-shopping across six websites, you tell your agent what you want and it does the work. Not for every purchase. People will still browse for fun, for discovery, for the experience. But for the utilitarian 80% of web interactions (pay this bill, find the cheapest option, schedule this appointment, send this follow-up), most people already do not enjoy this process, and delegation is structurally better.

Service Providers

This is counterintuitive, but service providers benefit enormously. On the human web, they spend billions on SEO, advertising, and UX optimization to win human attention. This has effectively made the internet a hostile place for humans, with companies constantly trying to get your attention and influence your behavior. In the agent internet, service providers declare their capabilities and compete on quality, reliability, and price. The most competent service provider wins, not the one with the best ad budget. Service providers may still want to influence agent behavior — economically beneficial and ethical ways to do this are outlined in From SEO to AEO.

Small service providers benefit disproportionately. A boutique retailer with excellent products but no marketing budget can declare their capabilities and receive agent traffic based purely on the quality of their service. The playing field flattens and the economy opens up.

Agent Developers

The agent internet creates a new developer ecosystem. Agent developers build specialized agents (travel agents, research agents, procurement agents, personal finance agents) that connect to the runtime layer and serve specific use cases. They're paid for orchestrating successful task completions, creating an economic category with no precedent in the human web.

The Vocabulary

To think clearly about the agent internet, we need shared terminology. I maintain a comprehensive glossary of terms used throughout this series.

What Comes Next

This paper establishes the frame. The subsequent papers in this series explore specific dimensions:

• The Three-Party Web — How trust, liability, and relationships change when an agent sits between user and service
• From SEO to AEO — How service provider discovery works in the agent internet
• Economics of Agent Commerce — How value flows when completion replaces attention
• Financial Services and the Agent Internet — Payment as the first and most critical meta-capability
• Identity and the Agent Internet — Attestation-based identity for agents acting on behalf of humans
• Communication and the Agent Internet — How agents speak, negotiate, and coordinate on your behalf
• Compute Efficiency as Value — Why efficiency is the new engagement metric
• Supervision as a Service — The trust and safety framework for agent autonomy
• Governance Through Architecture — Why the architecture is the governance
• The Agency Problem — Failure modes and safeguards for the agent internet

The agent internet is not coming. It's here. The question is whether it is intentionally built, with clear frameworks, fair economics, and human oversight, or we let it emerge chaotically.

I make a case for the former.