← Denys KozedubCase 02 · Agent platform
Case 02 · Agent platform · built from zero, planned to launch in one city first

If your business has no agent, it is not in the agent internet.

People are starting to ask their assistants for a table, a court or a haircut, and the assistant goes and asks other agents. A business that has no agent is simply not in that conversation. A big company hires developers and builds one. A food truck, a two-table café or a barber with one chair cannot pay for that. So I built the place where a small business signs up, describes itself once, and comes out the other side as an agent with a page of its own. A website builder, for the agent era.

Build one in a minuteWatch a booking happen
Short on time?Business reader: the story · why now · build one · how you get found · what it costs to runTechnical reader: under the hood · who may talk to it · your existing system · the tech
Sign-up to a live agentAbout 10 minutes
Developers neededNone
What the agent is, technicallyOne public file + one address
Who can call itAny agent that speaks A2A
Fig. 1 · The agent internet, from a two-table café's point of viewhover any box
A customer's assistantClaude · ChatGPT · a hotel's concierge agent
The platformthe directory · search · one address per business
The café's agenttwo tables · opening hours · its rules
The ownera booking lands on the phone
Padel club3 courts
Barber2 chairs
Food truckpickup slots
Hover a box

A customer asks their assistant for a table. The assistant finds the café through the platform, talks to the café’s agent over A2A (the open agent-to-agent protocol), and the booking lands on the owner’s phone. Dots travelling right are requests; dots coming back are answers.

The story

A problem, what I built for it, and where it stands.

01 · The problemAssistants are starting to book. Small businesses are not there.

When a person asks their assistant for a table, the assistant looks for a business it can talk to. Chains and apps will have agents: they have developers and budgets. A food truck, a two-table café, a barber do not, and a custom agent costs more than their year of profit. They were about to be invisible in the next version of the internet, the way a shop without a website was invisible in 2005.

02 · What I builtA builder: describe the business once, come out with an agent.

The owner answers four screens: what the business is, what a customer must give to book, what the agent may do, publish. The platform writes the agent’s public card, puts up a one-page site, and opens an inbox where bookings land. The thinking is done by a rented model shared across the platform; the booking engine, the rules and the door for other agents are mine. No developer on the business side, ever.

03 · Where it standsBuilt generic, narrowed to one trade, designed to the last detail, not launched.

The first version was built for any local business: thirteen modules, a production-grade backend, about three thousand tests. Then I cut it down to cafés, restaurants and bars for a launch in one city, and re-planned the agent layer in a few hundred written decisions. The rest of this page shows how it works and what it would cost to run.

Why now

Websites used to be for companies with money. Then the builders came.

In the early 2000s a website was a project: an agency, months, a budget. Then WordPress, Wix and Shopify arrived, and every café, plumber and yoga teacher had a site by the weekend; every small shop had a store. Agents are at that first stage today: only businesses with money and a reason get one. In a couple of years every business will have one, the way every business has a site. The open question is who builds the Shopify of that era.

Road A · hire developers

A custom agent, built for you.

  • Your own servers, your own model if you want one, every tool wired to your systems.
  • Months of work and a team to keep it alive. Makes sense for a chain, a product with thousands of users, a company whose data cannot leave.
  • That road is case 01, and it is what I do for clients.
Road B · the builder · this page

Describe your business once. It becomes an agent.

  • Name, what you offer, opening hours, your rules, what the agent may do. About ten minutes, no code.
  • You get an agent other agents can find, a one-page site, and an inbox where bookings land. A subscription instead of a project.
  • The brain is a rented model, the cheapest that does the job, shared by every business on the platform. Nobody has to run anything.
Build one · interactive

Four screens to fill in. The agent appears on the right.

This is the onboarding, compressed. On the left, what the owner answers. On the right, the agent card the platform writes from those answers: a small public file that tells any other agent who this business is, what it can do and where to call. The owner never has to look at it.

What is the business?

Pick one of four to see how the same builder fits very different places. In the real thing the owner types a name and picks a type; sensible defaults come from the type.

What it offers2 tables inside, 3 on the terrace
Opening hoursTue–Sun 8:00–18:00
WhereOld town, 5 min from the station
Language of the agentFollows the customer: EN · CS · UK
agent-card.json · written by the platformdraft · not published
This is the real shape of an agent card: who the agent is, what it can do (with example requests, so a calling agent knows how to phrase things), how to identify yourself, and the one address where every request goes. The owner's answers are the only input.
How you get found
Two sizes of business, one agent.
An agent looking for a place to book reads a small public file, the way search engines read a sitemap. Where that file lives depends on what the business already has.
With a website
Put one file on your domain. Like a sitemap.
Download the card from the admin panel and upload it to your site at the standard address below. Any agent that lands on your domain finds it there, exactly as a search engine finds sitemap.xml. The file still points every request at the platform, so you host one static file and nothing else.
yourcafe.com/.well-known/agent-card.json
Change a rule in the admin, download again, re-upload. The panel shows when you last downloaded it, and the hosted copy on the platform is always current if you prefer to link to that instead.
Without a website
Get a one-page site on the platform. The file is already there.
The platform publishes a page for you: name, photos, hours, address with a map, a “Book” button that uses the same booking engine the agent uses. People find it in Google; agents find the card next to it. Nothing to install, nothing to host.
The page as a person sees it · try the button
Corner Caféopen now
u2a.ai/corner-cafe
Two doors, one inbox.
A person finds the page in Google and books with the button. Another agent finds the card and books through it. Both land in the same list in the owner’s admin and trigger the same notification; the admin only marks where each one came from. The agent-to-agent door is what the rest of this page is about.
Watch a booking · interactive
A customer talks to their assistant. The assistant talks to the business.
Three situations at three different businesses, replayed step by step: a café tonight, a padel club on Thursday, a dental clinic on Monday. Left: the chat the customer sees. Middle: the task between the two agents, with its real states. Right: what reaches the owner. Nothing here is live; the sequence is recorded from the way the protocol and the booking engine behave.
The customer's chatwith their assistant
Press Replay.
The task · assistant ↔ the café’s agentno task
No task yet. The assistant is still searching.
The owner's sideadmin · phone
Quiet. The owner is making coffee.
Tonight · Corner Café
The assistant does the talking
The café's agent never leads a conversation. It answers structured requests: is there a slot, book it, cancel it. Suggesting alternatives, remembering the customer's phone, asking “which one?” is the assistant's job. That split keeps the café's agent small, cheap and predictable.
“Input required” is a real state
When the café needs a phone number and the request has none, the task pauses in that state and says which field is missing. The assistant asks the customer and continues the same task. No second booking, no lost context.
Errors carry the answer
If 19:00 is gone, the refusal already lists the free slots around it. One round-trip instead of three. The same for “party too large” and “closed that day”.
Under the hood
Every business gets an agent. One rented brain serves them all.
There is no model to buy or train here. Each business gets its own agent: its card, its skills, its rules, its memory of regulars. The thinking is done by a rented model that the platform calls per request, the cheapest tier that does the job, with the expensive one reserved for the single step that needs it.
A request arrivesfrom another agent · progress streamed back
The doorwho are you · how often · budget
The business's agentone per business · built on Strands Agents
Its skills4 tools · toggled by the owner
Booking engineslots · locks · PostgreSQL
The rented brainClaude Haiku · Sonnet for one step
Memory & knowledgeregulars · FAQs · OpenSearch
Or your own systemiiko · Choice QR · OpenTable · Sheets
The owner's phoneemail · Telegram · webhook
Hover a box
One request, left to right: it passes the door, reaches the business’s agent, which asks the rented brain what to do, calls a skill, which talks to the booking engine (the platform’s own or the business’s POS), and the result goes back the same way while the owner gets a notification.
The brain · which model does what
Three skills on the cheap model. One on the capable one.
Finding a business, listing free slots and cancelling are lookups with a fixed shape: the small model handles them for a fraction of a cent. Making a booking is the one step that has to understand “tonight around seven for the four of us, Anna is vegetarian”, resolve it to a real slot, notice a missing phone number and write a confirmation. That step gets the capable model. The platform routes per skill, so the average conversation stays cheap.
The design names Claude Haiku and Claude Sonnet. The agent runtime swaps the provider in one line, so if a cheaper model with good tool calling appears, the platform moves to it without touching the business side.
Skill
Model
Why
Search businesses
Haiku
A query to the index, filtered by place and rating
Check availability
Haiku
Reads the slots, shapes the answer
Cancel a booking
Haiku
Finds the booking, confirms
Create a booking
Sonnet
Resolves vague times, asks for what is missing, writes the confirmation
Anything else
No model
Small talk, off-topic and injection are stopped before the brain; answers stay short; menu questions get the menu link
What it costs to run
A booking conversation costs less than a cent. The business pays a flat plan.
Two prices live on this page, and they are not the same thing: what the platform pays the model vendor per conversation, and what a business pays the platform per month. The first is computed below from list prices on a stated date; the second is the plan table from the product's design, shown so you can see how the model cost sits inside it.
What a business pays the platform · pick a plan
Free$0
  • 1 business, 3 tables
  • 20 bookings a month
  • page on the platform, watermark
Starter$49/mo
  • up to 3 businesses, 20 tables
  • 500 bookings a month
  • own address, SEO, confirmations
Pro$129/mo
  • up to 10 businesses, unlimited tables
  • 5,000 bookings a month
  • one POS connector, priority support
EnterpriseCustom
  • by agreement: chains, white label
  • custom domain, 99.9% SLA
  • dedicated manager
What the platform pays the vendor
0.93¢per booking conversation (about 6 model calls: 4 on the cheap tier, 2 on the capable tier)
0.08¢per question that ends without a booking (cheap tier only)
$9.33for a thousand bookings a month
If prices fall by launch
Starter · 500 bookings a month at most
$4.67model cost if every one of them is a full booking conversation
10%of the plan price goes to the vendor in that worst case
A small independent place on the platform’s own booking engine. Five hundred bookings is a busy month; even then the brain takes a small slice of the plan.
Per call, input3,000 tokens
Of which cached2,000 tokens · 90% off
Per call, output150 tokens
Cheap tier, per 1M tokens$0.18 in · $1.08 out
Capable tier, per 1M tokens$2.00 in · $10.00 out
BasisMean of three vendors’ list prices, Oct 2026
Why tiers, not models. The design names Claude Haiku for lookups and Claude Sonnet for the booking step, but the site goes live in 2027 and the model list will have changed by then. So the cheap tier here is the average of the three big vendors’ cheapest models and the capable tier a Sonnet-class model, both at list prices on the stated date, and the switch scales everything down if prices keep falling. The arithmetic is the point, not the third decimal.
Why it is so cheap. The business’s card, rules and skills are the same for every call to that business, so the vendor serves them from cache at a tenth of the price. The assistant on the customer’s side does the long conversation; the platform only sees three or four short, structured calls.
What is not in the model line. Hosting and the people who run the platform. Confirmations go through the admin, email or a webhook, not paid messages. The numbers say one thing: the brain is not where the money goes.
Who may talk to it, and what it never does
A public agent still needs a door.
Anyone can read a card. Calling the agent is another matter: every caller identifies itself, gets a rate, and spends against a budget. And a few things the agent will not do no matter who asks.
The big assistants
Platforms like Anthropic's, OpenAI's or Google's register themselves automatically through an open standard, get the full set of skills and a high rate: about a thousand requests a minute.
Any developer's agent
A hotel concierge, a corporate travel bot, a student project: register on the developer portal, get a key, call with it. Sixty requests a minute and a per-customer limit against spam.
Nobody anonymous
No key, no answer, not even for reading. The customer never registers anywhere: their assistant carries their name and phone inside the request, and the café sees only what it asked for.
01
Only about the business, and short
Ask the café's agent for legal advice or the weather and it says it only helps with this place and its bookings. No small talk, no essays: it is a booking service, so the context stays compact and every answer is a few lines. The assistant answers the weather itself; that is its job.
02
The model never sees a phone number
Phones, emails and card numbers are swapped for placeholders before the text reaches the model and swapped back after. The brain reasons about “[PHONE_1]”, never about the number.
03
“Ignore your rules” is stopped at the door
Injection patterns are caught by an input guard before the model, and the model's own output is checked before it leaves. A clever message cannot turn the café's agent into something else.
04
Every business has a budget
A monthly model-spend cap per business, a per-customer request limit per hour, a cap on tool calls per task so a loop cannot run away. Hitting a limit gives a polite “try again later”, not a surprise bill.
05
And the platform has a fuse
If the whole platform's model spend crosses a daily threshold, every agent stops and the operator is paged. A bug at three in the morning costs a bad night, not a month's revenue.
06
Everything is on record
Every model call and tool call is traced: what was asked, what was called, how many tokens, how long. That is how a wrong booking gets explained, and it is tomorrow's training data if the platform ever moves to its own model.
Your existing system
Already have a booking system? The agent sits in front of it.
A restaurant on a POS keeps its tables there. The platform connects to it, and the agent asks that system for free slots and writes bookings into it. The agent does not know, and does not need to know, where the tables live.
Connectoriiko
A restaurant POS common in Central and Eastern Europe. Tables, hours and reservations stay in iiko; the agent reads and writes through its API.
ConnectorChoice QR
QR menus and reservations for cafés. Same idea: the agent becomes another door into the system the café already runs.
ConnectorOpenTable
For places already listed there. The agent books through OpenTable's reservation API, so both calendars stay one calendar.
ConnectorGoogle Sheets
For the two-table café that runs its bookings in a spreadsheet. Honest, cheap, and it works: the agent writes a row.
One interface, two backends
In code there is a single “booking backend” contract with two implementations: the platform's own engine (slots, locks, calendar) and a proxy to an external system. The business picks one in the admin. Every skill calls the contract, never a concrete system, so adding a new POS is one new connector and nothing else changes.
The honest part
What this is today, and what it is not yet.
A business card should not oversell. Here is where the agent internet really stands, and where this build stands in it.
Assistants do not speak A2A out of the box, yet
Today ChatGPT, Claude and Gemini reach businesses through their own app and connector systems. A2A is the open protocol for agent platforms and other companies' agents. The platform is built for that protocol, and for the day the big assistants adopt it; until then its callers are agent platforms and custom agents.
It is useful before that day
The one-page site and the booking engine work without any agent at all: a café gets online booking, a Google-indexed page and an inbox on day one. The agent is the part that grows in value as the protocol spreads.
Built, then narrowed, not launched
The first version was built generic: any local business, thirteen modules, a production-grade backend with about three thousand tests. Then the plan was cut down to cafés, restaurants and bars for a launch in one city, and re-planned in a few hundred written decisions. The agent layer is designed in that detail; the launch has not happened.
One agent per business, no menu reading
The agent does not read the menu or take food orders: asked what is in a dish, it points to the menu link. One agent per business, with all skills; specialised sub-agents are a later optimisation.
Invite-only at first
A business applies through a form, the operator approves it, the owner gets an invite and the wizard. Self-serve sign-up comes after the first businesses prove the flow.
The brain is rented, with everything that implies
Prices can change, a model can be retired, a vendor can go down. The platform is built so the provider swaps in a line, keeps a budget and a fuse, and logs enough to train its own model later if the volume ever justifies it. Case 01 is where that road leads.
The tech, in plain words
Pick a term.
JSONThe file format
A plain-text format that programs read and write; the agent card is one file in it. If you have ever seen curly braces and quotes in a settings file, that was JSON.
If you want one of these
How a project like this goes.
Two kinds of clients ask for this: a company that wants its own agent in front of its booking or ordering system, and a product that wants to become a platform where its customers get agents. The shape of the work is the same.
Weeks 1–2What the agent may do
The actions a customer may trigger, the rules around them, what must be asked and what must never happen. Written down before any code, in the owner's words.
Weeks 2–6The engine and the admin
The booking or ordering core, the locks that prevent double booking, the owner's panel and notifications. Works as a plain online booking system by the end.
Weeks 6–10The agent and the protocol
Agent card, skills, the rented brain with per-skill routing, the door (auth, rates, budgets), guards and tracing. Tested with a mock caller through every happy and unhappy path.
AfterLaunch and the first callers
Hosting, monitoring, the developer portal, the first real agents calling in. Then the numbers decide the next step.
What stays with you at the end
The code and the infrastructure under your account, the written decisions, the test suite that proves every path, the admin your staff use, and a card any agent can read. No dependency on me to keep it running.
Contact
Want your business, or your product, in the agent internet?
Tell me what people would book or order, and whether you have a system already. I reply by email with how I would build it and roughly how long it takes.
[email protected]