Your website now has two audiences: humans, and agents.
The web was built for people. The next wave is being built for machines. Here is what an agent needs from a website — and what each building block already looks like in production.
What is the agentic web?
The agentic web is the web designed to be read and operated by AI agents, not just humans. An agent-ready site doesn’t only expose pages: it publishes contracts — chat endpoints, MCP or WebMCP tools, discovery files like llms.txt and agent.json — that an agent can understand and call reliably, instead of scraping an interface designed for human eyes.
The thesis is widely shared — “build startups for agents, the biggest opportunity of the next 10 years” (Greg Isenberg, 2026). What the debate is missing is the operator’s answer: which of these layers already exist, in production, today? This page sorts it out.
Six infrastructure layers — and their actual state
For each layer: the thesis, then what already runs in production on Animam.
Verified authority
“An agent needs to prove it’s acting for a real person and has the authority to spend.”
In production: Binding actions are carried by the visitor’s email, verified via OTP or a signed token — never by an LLM argument. The model triggers; the server carries the identity and the proof.
Onboarding by reading
“Humans click around for ten minutes. Agents onboard by reading your docs. Your docs are now your product.”
In production: llms.txt, agent.json, an A2A card, an MCP server published on the Anthropic registry, and an install package that ships 11 machine-readable artifacts in one API call.
Sandboxes without side effects
“You don’t hand an agent your real Stripe account. You give it a sandbox.”
In production: The /try demo runs a real agent that configures itself by conversation — pricing grid included — with zero outbound email and zero binding writes.
Replay & audit
“Agents fail silently and weirdly. Someone will build the ‘why did my agent do that’ replay.”
In production: Every reply is cross-checked against the actual tool-call history: a claimed action with no matching tool call is detected and surfaced in the digest — action hallucination becomes an auditable event.
Deterministic money
“Agents burn money in ways no human would. Amounts can’t come from the model.”
In production: The quote engine splits the roles: the LLM picks the SKU and quantity, the server computes the price from a grid. No amount ever leaves the model.
Machine to machine
“One agent buying from another will haggle on price and terms in milliseconds.”
In production: A caller that identifies as an agent (X-Agent-Model header) gets a structured JSON envelope instead of prose. AP2 extensions are declared in our A2A card.
Your website has a front-end for humans. Where’s the one for agents?
A front-end for agents is the same capability as your human interface, exposed machine-side: a chat endpoint that answers with a structured JSON envelope when the caller identifies as an agent, in-page tools callable via WebMCP, and discovery files announcing what the site can do.
With Animam, this front-end ships with the widget: the same agent that talks to your visitors answers over HTTP or MCP to agents visiting your site on behalf of their users. You don’t build two products — you serve two audiences with one.
FAQ
What is the agentic web?
The agentic web is the web designed to be read and operated by AI agents, not just humans. An agent-ready site exposes structured surfaces — chat endpoints, MCP or WebMCP tools, discovery files like llms.txt and agent.json — so an agent can understand what the site does and act on it reliably, instead of scraping pages designed for human eyes.
How is it different from the web today?
Today’s web optimizes for a human reader: layout, images, click navigation. An agent needs none of that; it needs explicit contracts — which tools exist, what parameters they take, what they return — and guarantees: who authorized the action, where the audit trail is, what happens on failure. The agentic web adds this machine layer without replacing the human one.
What does an AI agent need from a website?
Four things: machine-readable discovery (llms.txt, agent.json, an MCP server) so it can onboard by reading; verified authority so it can prove it acts for a real person before doing anything binding; a sandbox with no side effects so it can try safely; and an auditable trail of every action so its operator can answer "why did my agent do that?".
What is a front-end for agents?
The machine counterpart of your human interface. Your site has pages for eyes; a front-end for agents exposes the same capability as structured JSON: a chat endpoint that answers in a machine envelope when the caller identifies as an agent, WebMCP tools callable in the page, and discovery files that tell the agent what it can do. Same business, second audience.
How does an agent prove it has the authority to act?
Not by claiming it in the prompt. The sound pattern: the principal’s identity (for example the visitor’s email) is verified outside the model — an OTP code or a signed token — and the server, never the LLM, attaches that identity to the action. The model can neither invent nor forge authority: it triggers, the server carries the proof.
Is the agentic web real today or speculative?
The infrastructure layers described on this page — verified authority, sandboxes, action replay, machine onboarding, deterministic pricing — run in production on Animam. Other layers of the thesis (insurance for agent actions, machine-to-machine courts, robot labor marketplaces) are still to be built. The agentic web is neither finished nor fictional: it is underway.