Seven agents. One walk through your store.

Three walkers browse like your buyers and capture everything. Three analysts reread the walk, one mandate each. A judge tests your filters against named personas. All seven run on every audit.

Public pages only. No integration, no card.

RUN 041 · thequest.ch PERSONA: PRESSED-FOR-TIME PARENT
07:00:04walkernavigate /collections/chaussures capture ✓
07:01:37walkersearch "speedgoat 5" → 0 results capture ✓
07:04:12walkerproduct ×2, cart, checkout step 1 capture ✓
07:07:39walkerwalk closed — 14 pages, judged nothing
07:07:41analyst·trustdelivery time absent at point of doubt
07:07:52analyst·wayfinding"speedgoat 5": product exists, search finds nothing
07:08:15analyst·funnelaccount required, reason never given
07:08:20tower7 findings, each with its capture → review
end-to-end 455 s pages 14 findings 7 generated proof none

Nobody grades their own walk.

Walking the store and judging it are separate jobs, held by separate agents. That split is what keeps the findings honest — and it is enforced, not encouraged.

Walk

A walker browses as your persona: category, search, product pages, cart. Every page it opens is captured.

walkers never judge

Reread

Analysts read the dossier the walker left. Each declares findings in its own categories; anything outside its mandate is refused.

analysts never browse

Review

Findings land in your inbox with capture and path. You judge each one — actionable, false, noise — and your verdicts train the next run.

tower never self-approves

The fleet.

Each agent below is a prompt file in the repository, quoted here as written. All seven are available on every audit.

Walkers

×3 They open, read and capture. Their prompts start with the same order: judge nothing.

  • Generalist walker

    runner/src/prompts/parcours.md

    Buys like a person with real intent. Types at least two search queries and opens at least two product pages — a supplier-copied description only shows itself when you compare.

    route: home → category → search → product ×2 → cart
    persona: injected per run

  • Search walker

    runner/src/prompts/parcours-recherche.md

    Spends its entire budget on your search engine. Runs neighbouring queries against each other: the same product found under one name, missing under another; a count that contradicts the list.

    surface: search only
    method: query classes, compared in pairs

  • Filter walker

    runner/src/prompts/parcours-filtres.md

    One surface: the filter panel on category pages. Unfolds every group, reads every value, records the default state — a folded group the walker skips will not exist for anyone downstream.

    surface: category filters only
    budget: all of it, on the panel

Analysts

×3 They read the dossier and cannot return to the site. Overlapping mandates are refused at the door.

  • Trust analyst

    runner/src/prompts/analyse-confiance.md

    Returns, delivery, stock, payment, contact. Its question is never does the information exist somewhere — it is is it there where the doubt is born. A delivery time in the footer is not on the product page.

    declares only: reassurance · proposition_valeur
    site access: none

  • Wayfinding analyst

    runner/src/prompts/analyse-reperage.md

    Knowing where you are, going back, comparing. A category unreachable from home, a false breadcrumb, a search that returns nothing while the product exists, a back button that loses your state.

    declares only: navigation · friction_ui
    site access: none

  • Funnel analyst

    runner/src/prompts/analyse-tunnel.md

    Every step of checkout that demands effort without explaining it: a field that asks why-less questions, an account wall with no stated reason, a surprise waiting at the cart.

    declares only: tunnel
    site access: none

The judge

×1 Specialised verdicts on one surface, verified from outside.

CRO Filter Judge

design/agents/cro-filter-judge.md

Takes one category page and two named personas with contrasting decision drivers — a comfort-first novice against a spec-literate expert. Says whether your filters let each of them find what they came for without guessing or opening every product. It renders findings in prose, never a score.

inputscategory · 2 personas · panel capture
outputfindings in prose, no score
grid5 pieces, verified independently
verified againstBaymard Institute, blind

Measured, not promised.

Every number on this page has a run log behind it. That includes the ones we got wrong first.

Full walk, end to end 455 s → 7 findings run log 06-05
Findings judged one by one by the store owner 11/11 true pilot review 06-16
Store coverage, first audit → second audit 65 % → 90 % run log 06-16
Retrieval misses before the fix — published, not buried 43/70 eval report

Send the fleet through your store.

First audit free. You judge every finding.