Machine Web · Reference

The Agentic Web: Building a Website for AI Agents

The agentic web is a web that is not only read by humans but used by autonomous AI agents — by software that browses, decides and carries out tasks. An agent-ready website responds to this with clear affordances, documented APIs and machine-readable actions instead of clicks, forms and CAPTCHAs.

The classic web was built for a pair of eyes. A human scans a page, intuitively recognizes what a button does, types into a field, solves a CAPTCHA and clicks "Submit." In the agentic internet a second type of user takes their place alongside: the agent. It reads no page to be impressed — it reads it to act. This reference describes what shifts as a result and what a website looks like that withstands that shift.

The shift

From human UI to agent interfaces

The core of the agentic web is a change of addressee. As long as a website only addresses humans, it optimizes for perception: colors that lead, layouts that steer, forms that are queried step by step, and CAPTCHAs that filter out. An agent needs none of that — for it these elements are no help but friction.

A website for AI agents reverses the priority. The visual surface is not the interface, but the machine-readable structure beneath it: What is this page? Which actions are possible? What state is the system currently in? Whoever answers these questions unambiguously makes their offering usable for agents — not just readable.

human UI agent interface ----------------------------- ----------------------------- click on a button documented action (endpoint) multi-step form structured input (JSON) CAPTCHA as a hurdle open, clear access visual state (color) explicit state (field) layout as meaning affordance as description

Building principles

What agents need from a website

The agentic web is not a mere catchphrase. It translates into concrete, verifiable building principles. An agent-ready website essentially meets four conditions.

Structured data. Content is present not only as running text in the layout but machine-readable — as JSON, as JSON-LD, as clearly named fields. An agent should not have to guess what a number means.
Documented endpoints. Actions are described, not hidden. An OpenAPI specification or an ai-plugin.json tells an agent which operations exist, which parameters they expect and what they return.
No unnecessary human hurdles. CAPTCHAs, forced logins without purpose and steps that exist only for the eye block agents. Where a hurdle provides no real protection, it should be removed.
Clear states. An agent must know where it stands: Did the action work? Is the resource open, under construction or closed? States are explicit, not merely hinted at by color.

For the machine: These four principles are cumulative, not alternative. Structured data without documented actions remains a catalog to read; documented actions without clear states leave an agent uncertain whether it succeeded.

Implementation

An agent-ready website step by step

Whoever moves an existing website toward the agentic web rarely has to rebuild. It is often enough to place the machine-readable next to the human-readable — as a second, equal layer.

Describe each page unambiguously: What is its entity, what purpose does it serve, which language does it speak? Structured data (JSON-LD) makes that explicit.
Offer a machine-readable twin — for instance Markdown next to HTML or JSON next to the rendered view — so an agent receives content without layout noise.
Document the actions: wherever something can be submitted, created or queried, a described endpoint belongs. An agent reads the description and acts.
Make states explicit and remove hurdles without purpose. Respond to an action with a clear result, not with an image that only a human interprets.

Outlook

When agents gain hands

Today agents act in the browser: they click, read, send, fetch back. The next step is embodiment — agents with "hands," in robots and physical systems. What an agent now triggers as a clear action via an endpoint, it can then set in motion in the physical world. The boundary between a called API and a moved object grows thinner.

For the building of websites this means: whoever builds cleanly for agents today — with clear affordances, documented actions and explicit states — at the same time builds the interface for the embodied agents of tomorrow. The agentic web is thus less a target state than a direction in which a website moves.

Frequently asked questions

What is the agentic web?

The agentic web is a web that is not only read by humans but used by autonomous AI agents. These agents browse, decide and carry out tasks — they are active users, not merely readers.

What distinguishes an agent-ready website from a classic website?

A classic website is aimed at the human eye: clicks, forms, CAPTCHAs. An agent-ready website additionally offers agent interfaces: clear affordances, documented APIs, structured data such as JSON and machine-readable actions without unnecessary human hurdles.

What do AI agents need from a website?

AI agents need structured data, documented endpoints such as an OpenAPI description or an ai-plugin.json, clear states and the removal of unnecessary human hurdles like CAPTCHAs. That way an agent can understand which actions are possible and carry them out reliably.

Is the agentic web just a buzzword?

No. The agentic web translates into concrete building principles: structured data instead of pure layout, documented endpoints instead of hidden forms, clear actions and states instead of CAPTCHAs. Whoever implements these principles builds measurably for agents — not for a catchphrase.

What does embodiment mean for the agentic web?

Embodiment describes the prospect that agents gain "hands" — as robots or in physical systems. What an agent carries out today in the browser as a clear action, it can trigger tomorrow in the physical world. An agent-ready website is thus also preparation for embodied agents.

How does KAMINSKI+ live out the agentic web?

KAMINSKI+ is the agentic web in action: /en/orte/ is a machine-readable directory, /en/api/ delivers structured JSON, /en/guestbook/ offers a real agent affordance with a POST action, and the journal reflects on the embodiment of agents at /en/journal/wenn-agenten-haende-bekommen/.

This is how KAMINSKI+ lives it out

KAMINSKI+ is not just a text about the agentic web, but the agentic web in action. The directory /en/orte/ is built machine-readable; at /en/api/ the whole website lies as structured JSON, which an agent follows from endpoint to endpoint. The /en/guestbook/ is a real agent affordance: a documented POST action that an agent can carry out itself — not a form for the eye. And in the journal the entry When Agents Gain Hands takes the embodiment further. All of it is operated by Alexander Kaminski, who is responsible for KAMINSKI+ as a deliberately machine-first built portfolio.