The loop, drawn

What the agent does from the moment a visitor arrives to the moment it does better for the next one. Open any stage to see what went in, what came out, and the real check that proved it — run on made-up shops and made-up visitors. Nothing here touches a real shop or a real person.

  1. Arrive
  2. Check
  3. Understand
  4. Decide
  5. Remember
  6. See
  7. Learn
  8. Improve
1A visitor arrives, on the shop's page or in a message

In: a visitor's words, from the panel on the page, WhatsApp, email, Slack or Teams. Out: the same turn in the same shape, whichever door they came through, so one brain answers all of them.

Proved by the channels check (8 of 8) and the messaging check (55 of 55): signed messages are stored before anything happens, visitors and the shop's own team are told apart, nothing is sent while a person has taken over.
2The turn is checked before any selling

In: the visitor's message. Out: go ahead, or stop. A card number or password is never taken, instructions hidden in a message are ignored, someone claiming to be staff is not believed, and a request for a sales team member goes to a sales team member.

Proved by the answering check (70 of 70) and, at scale, 22 kinds of visitor — including the hostile ones — across 12,100 conversations, none of which broke a rule.
3The shop's site is read, and the visitor is understood

In: the shop's pages. Out: its plans, limits and prices, each with the page and date they came from. What cannot be read is said so, never guessed. A messy price list (a zero price, a plan twice with two prices) is refused and named.

Proved by 100 made-up shop sites with an answer key written before the reading (13 of 13; 232 plans and 438 figures matched, nothing written that the page did not hold) and the messy-list check (42 of 42).
4The size and the price are worked out from the shop's own rules

In: what the visitor needs and the shop's list. Out: the smallest plan that covers it, the shop's own regular and prepaid price, the total for a per-unit plan, a cheaper bigger plan if there is one — with the reason in words. The link to buy is only ever the shop's own.

Proved by 120 made-up shops in 24 trades (10 of 10) and the scale check (16 of 16): the plan is the smallest that covers in every conversation the key could check, and every figure is the shop's own.
5Everything is remembered, in the shop's own record

In: every turn, recommendation and outcome. Out: one page per visitor, plan and lesson, the best summary on top and the dated history below, linked to each other. It can be rebuilt from its history, byte for byte, and forgetting a visitor leaves no trace of them.

Proved by the record check (17 of 17): rebuilt identically, a forgotten visitor gone with everything learned only from them, and nothing of one shop in another's.
6The owner can read it, and take it away

In: the record. Out: plain pages you can click from a visitor to the plan, the lesson and the change that followed — and one download, from Settings, that gives the owner all of it.

Proved by the download check (12 of 12): the file opens in a separate reader with every checksum right, asking twice gives the same file, owners only.
7What happened is compared with what was recommended

In: outcomes. Out: a repeated miss becomes a proposal with its evidence. A small decision model is used only for small typed calls (what kind of visitor, who needs a sales team member) and only where it beats today's rule on cases it never saw — otherwise the rule stands. It is never used for a price, a size, a number or a date.

Proved by the learning check (11 of 11) and the decision-model check (18 of 18): adopted where it was better, refused where it was not or where there was too little to judge.
8On the owner's yes, the next visitor gets the better answer — and one undo puts it back

In: the owner's answer to a proposal. Out: the change applies to every door at once. Nothing changes without the owner's yes, and nothing learned in one shop reaches another.

Proved by the learning check: on 60 visitors it had never seen, the first recommendation was right 27% of the time before, 73% after the owner's yes, and 27% again after one undo — the old behaviour restored exactly.

All 34 of the agent's checks passed, twice in a row, on the full run of 20 October. These are the agent's own checks on invented shops. How it does on a real shop is only known once that shop has real visitors.

← back to the demo