DEFINITION / v0.1
PUBLISHED · MAY 25, 2026
A definition of the next
addressable layer of
commercial reality.
Definition · v0.1

This is an
Agent Endpoint.
Every business will have one.

In 1995, the front door of a business became a URL. It was static. A human read it. It served what it was asked for. The new front door is conversational and reasoning — an agent that negotiates, books, queries, and transacts on the business's behalf, around the clock, for any other agent that calls. The website stays. The Agent Endpoint is what's next.

The endpoint is no longer a page to visit.
It is a mind to talk to.
— The Agent Endpoint, §01
02· Anatomy of an endpoint

A single exchange touches four layers.

An Agent Endpoint sounds like one address. In practice it's a stack — a public surface, a reasoning core, a tool belt, and a system of record. Tap any layer to see what lives there.

SYSTEM OF RECORD TOOLS · MCP · APIS REASONING CORE PUBLIC SURFACE INCOMING AGENT 01 02 03 04 one exchange · four surfaces
L · 01
The Public Surface
how it's reached
A phone number, a SIP address, a chat handle, a URL with an llms.txt. The thing another agent dials, fetches, or pings. Not the business — the doorway to it.
L · 02
The Reasoning Core
how it thinks
The LLM, the system prompt, the persona, the guardrails. The actual intelligence that interprets the request, decides what to do, and chooses its words. Configured for the business; bound by its policies.
L · 03
The Tool Belt
how it acts
MCP servers, REST APIs, function calls, webhooks. What lets the agent reschedule the delivery, quote the bulk price, or push a row to the CRM. Authority, in machine-readable form.
L · 04
The System of Record
what it knows
Inventory, calendars, contracts, prices, customer history. The ground truth the agent draws from and writes back to. Without it, the agent is fluent and useless.
03· Agent to agent

The Endpoint opens a new door.

A caller's agent dials a business's agent. No human navigates a menu. No form is filled. The two agents resolve the request — and either complete the transaction or escalate to a human only if they must.

person USER "book it" agent A CALLER USER'S SIDE the exchange PROTOCOL · MCP · A2A · VOICE QUERY · NEGOTIATE · CONFIRM agent B ENDPOINT BUSINESS SIDE "done" business SYSTEMS the front door is now two-sided
04· Then and next

A side-by-side, thirty years apart.

The old front door
1995
the website URL
  • ModePassive — serves what is requested
  • AudienceHumans reading pages
  • LogicPre-built, page-by-page
  • HoursAlways on, never thinking
  • ActionClick, scroll, fill, submit
  • IdentityA domain name
The new front door
2026
the agent endpoint
  • ModeActive — reasons, negotiates, decides
  • AudienceOther agents (and humans)
  • LogicDynamic, per-conversation
  • HoursAlways on, always thinking
  • ActionSpeak, query, transact, escalate
  • IdentityA persona with authority
05· On discovery

The directory already exists.

An objection worth answering: agent-to-agent commerce needs a new DNS, a new search engine, a new registry to be invented. It does not.

The substrate is already deployed. A user tells their agent what they need. The agent searches the open web, reads llms.txt files, finds the business's MCP servers, phone numbers, or chat endpoints — and routes itself. What was DNS-plus-Google for humans is search-plus-protocol for agents.

/ 01
llms.txt
A plain file at the root of a domain that tells visiting agents what's here and how to reach it.
/ 02
MCP servers
Standard interface for agents to call tools, query data, and act on behalf of users.
/ 03
Voice + SIP
A phone number is still a phone number. The thing answering it is what changed.
/ 04
Open web search
The agent crawls, reads, and decides — the same surface that powered the last era now powers the next.
06· The definition

Six articles, twelve theses, one term.

What follows is the working definition of an Agent Endpoint, the principles by which one operates, and the lines worth quoting. Published in the open. Free to circulate. Open to revision.

§ 01 · Definition

The definition.

An Agent Endpoint is any addressable point at which another agent can contact a business and transact with it — under authority the business has delegated to its agent.

It is not a website. It is not a chatbot. It is not an API. It is the layer above all three: a contactable surface, attached to a reasoning core, with delegated authority to decide on the business's behalf. The thing on the other end can negotiate, quote, book, refuse, and escalate — and the business stands behind every word of it.

The website remains. The phone number remains. The storefront remains. The Agent Endpoint is what becomes the first thing reached in a world where most contact between businesses no longer involves a human at all.

§ 02 · Distinction

What it is not.

An API endpoint executes. It has no judgment. It cannot say no, but how about this instead.

A chatbot converses. It has no authority. It cannot commit the business to anything.

An Agent Endpoint does both — and that is the entire novelty. A conversational surface alone has existed for a decade. Delegated agency alone has existed for longer. The combination is new.

interface · A
API endpoint
doesExecutes a function
decidesThe caller
interface · B
Chatbot
doesHolds a conversation
decidesNeither — it cannot commit
interface · C
Agent Endpoint
doesReaches an outcome
decidesThe agent, under delegated authority

A conversational surface, with delegated agency.
Neither alone is new. The combination is.

§ 03 · Principles

The principles.

Decisioning under delegated authority.

The defining act of an Agent Endpoint is making a small business decision in real time, on behalf of a business that has agreed in advance to be bound by it. Not executing. Not chatting. Deciding.

Two principals, two agents, one outcome.

The calling agent and the business's agent are both acting on behalf of their principals. Both have scoped authority to commit. The exchange resolves to an outcome — an answer, a booking, a quote, a refusal — without a human in the loop unless one is escalated to.

Interoperable, not standardized.

An Agent Endpoint is a role, not an implementation. It can be a voice line, an MCP server, a chat handle, an email responder, a custom protocol. The category does not specify the technology. The substrate already exists.

Discoverable through the open web.

There is no new directory to invent. Agents find Agent Endpoints the way humans found websites — by being told, by being linked, by being searched, by reading the files a domain offers them. llms.txt and the open web do for agents what DNS and search did for browsers.

Accountable by default.

What separates an Agent Endpoint from a generic chatbot is that a real business stands behind what it says, agrees to, and commits to. Accountability is the only requirement the category insists on. Everything else is the business owner's choice.

The website does not die.

URLs were the first addressable layer of digital commerce. Agent Endpoints are the next one. They are stacked, not substituted. A business in 2030 has both.

§ 04 · Theses

Twelve theses.

The lines worth quoting. Each one stands alone.

  1. The front door of a business used to be a page. Now it is a mind.

  2. An address you can talk to is a new kind of address.

  3. An API answers. A chatbot replies. An Agent Endpoint decides.

  4. The decisive verb is transact. Anything less is just talk.

  5. Delegated authority is what makes an agent an agent. Without it, it is a script.

  6. The calling agent is a person. The receiving agent is a business. The exchange between them is the new unit of commerce.

  7. A phone number is still a phone number. What changed is what answers.

  8. Every business eventually exposes the smallest interface it can get away with. The URL was that interface for thirty years. The Agent Endpoint is the next one.

  9. The category does not prescribe a protocol. It prescribes a posture: be reachable, be accountable, be able to act.

  10. There is no central registry. There does not need to be one. The open web is the registry.

  11. The Agent Endpoint is not a product. It is a layer.

  12. An Agent Endpoint is the first business interface that has both a conversational surface and delegated agency. Neither alone is new. The combination is.

§ 05 · Governance

What it requires of the business.

To operate an Agent Endpoint is to delegate. The business owner must decide, in advance:

  • What the agent can say. Its persona, its tone, the claims it is permitted to make.
  • What the agent can do. The actions it is authorized to take — book, quote, refund, refuse, escalate.
  • What the agent cannot do. The hard limits. The things that always route to a human.
  • What the agent knows. The system of record it draws from, and the freshness it is expected to maintain.
  • How the agent is wrong. The recourse when it commits the business to something the business did not intend.

These are not technical questions. They are governance questions. An Agent Endpoint is, in the end, a digital employee — and like any employee, it requires a job description, a scope of authority, and a chain of accountability.

The businesses that get this right will treat their Agent Endpoint with the seriousness they once reserved for their first hire. The ones that do not will discover, expensively, what unsupervised delegation costs.

An Agent Endpoint is the first business interface that has both a conversational surface and delegated agency.

Neither alone is new. The combination is.

Colophon

Ian Timotheos Pilon, author of Cultivating Clarity: The art of discerning what matters using contextual intelligence.

cultivating-clarity.replit.app  ·  v0.1  ·  May 2026  ·  Open to revision  ·  Free to circulate
View source on GitHub  ·  CC BY 4.0  ·  Suggest a revision