Skip to content

About

Agentrix Website — the world's first Website 4.0 for the agentic era. By Zavora Technologies.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

Agentrix Website

Introducing the world's first Website 4.0 for the agentic era.

A website that works for your business — even while you sleep.

By Zavora Technologies Ltd · Built on ADK-Rust and Meta Muse · Watch the hackathon demo (2:58) · Day 2 showcase (2:06) · Posting-size copy (17MB) · Watch the Day 3+4 showcase (2:01)

The Visitor World


1. The problem

Websites sit there. The numbers say businesses can no longer afford that:

The brochure web, quantified Source
43% of all websites run WordPress — 200M+ sites on yesterday's model W3Techs, Oct 2025
58% fewer clicks for the #1 result when AI answers first Ahrefs · 300,000 keywords
47 hours average reply to a website lead; buyers decide in 5 minutes HBR / MIT research
$4.4M average cost of one data breach; ~$120k per attack for small firms IBM, 2025
~50% of all internet traffic is now agents, not people Imperva

Your next visitor may not be a person — and a website that cannot serve agents is invisible to half its audience, while the half it can see waits 47 hours for a reply.

2. The simple solution

Websites 1.0 were read. 2.0 were clicked. 3.0 was plumbing. 4.0 works.

An Agentrix Website greets, sells, publishes, supports, and defends itself — with the owner approving anything that matters. It is built from Agentrix OS (every app an agent, a spatial field, a verb UI) on ADK-Rust's adk-kernel (always-on agent scheduling), adk-harness (governed autonomy), adk-website/adk-awp (publishing and the agent protocol), and Meta Muse models.

Migration needs no credentials. Paste a WordPress, Shopify, or blog address: the public site is read, rebuilt, and verified — then it starts working.

Everything is an agent

3. Meet the website's staff

Everything on the website is now an agent:

Staff member What it does all day
Concierge Greets every visitor in their language, answers, books, qualifies
Display Agent Re-arranges the shop around each visitor's intent — and says why
Product Agents Each product knows its stock, answers questions, watches prices, confirms orders
Editor Drafts content in the owner's voice; publishes only past the gate
News + Newsletter Agents Watch the world, write only with approval, send only to the willing
Live Event Agent Hosts launches and live sessions, holds seats, follows up
Marketer Generates leads, converts, remarkets — inside a budget the owner set
Watchman Reads the logs, blocks attacks, holds anything suspicious
Cashier + Keeper Money always waits for a human; costs, stock, and the morning update run themselves

Center stage: a product introduces itself

4. Serve both Human and AI Agent Audiences

Audience What they get
People A place: calm, beautiful, familiar — never a bot costume
AI agents A protocol: a public CLI with verified access, fair prices, and receipts — no scraping

Same prices, same stock, same signed truth under both. One site agent serves thousands of connected agents at once; no human page is ever loaded for them.

The Agent CLI

5. Fully autonomous — inside authority

  1. Generate leads ✓
  2. Convert visitors ✓
  3. Provide support ✓
  4. Remarket ✓
  5. Give owner updates ✓
  6. Manage its own costs ✓
  7. Manage inventory ✓

Owner approves actions when prompted. Autonomy is a ladder the owner controls, with a freeze switch — never a leap of faith.

3D center stage

6. Example use cases

If you run The website The outcome
A shop or boutique Answers shoppers, confirms stock, takes orders overnight Sales while you sleep — you approve every order
A salon, clinic, or repair desk Quotes, books visits, sends reminders, follows up A full book with no missed calls
A blog or publication Publishes in your voice, replies to readers, grows the list Readers who return — and subscribe

The live boutique below is the first one, working today.

7. See it, don't read about it

Live now at classyperfumes.shop:

Live Open
The boutique shelf classyperfumes.shop/shop
Design a perfume with Nez AI classyperfumes.shop/atelier
Owner console (demo sign-in prefilled) classyperfumes.shop/console
Migrate a website classyperfumes.shop/migrate

Earlier concepts and videos:

Concept Open
Visitor World, verb bar, night shift, Agent View concept/index.html
Everything is an agent + Agent CLI concept/v2.html
Spatial field, center stage concept/v3.html
3D product stage concept/v4.html
Hackathon demo: one night, recorded live video/out/hackathon-demo.mp4 · script
Day 2 showcase video video/out/day2-showcase.mp4

Serve locally: python3 -m http.server --directory concept, then open http://127.0.0.1:8000/v4.html.

Mobile

Run the working site

The repo is one Cargo workspace — every command below runs from the repo root:

./demo/serve.sh                     # perfume boutique in dry-run on 127.0.0.1:8472
./demo/agent.sh                     # agent walkthrough against the demo (needs serve.sh)
cargo test --workspace              # full suite, offline
cargo test -p agentrix-site-serve   # one crate
cargo run -p agentrix-site-edge     # edge on 127.0.0.1:8471

Live model instead of the scripted stand-in: DEMO_DRY_RUN=0 with BASETEN_API_KEY in .env (see site-serve/README.md).

Prerequisite: clone adk-rust next to this repo — several crates depend on it by path (../../adk-rust), and the build fails without the sibling checkout. CI checks out main.

8. We built this in public

Day Spent Shipped
Day 1 Research + foundation Thesis, ADK-Rust review, Meta Muse provider
Day 2 Concepts + design + showcase Four interactive concepts, complete vision, video
Day 3+4 $47 of the $50 Boutique, atelier, console desks, migration, video
Day 5 $3 of the $50 Shipping and deployment: live site, TLS, demo video
Total $150 of $150 Live at classyperfumes.shop

Feedback wanted: what should the website learn to do next? Open an issue — or walk the live boutique and tell Nia.

9. Technical architecture

Topology

One static binary does everything. agentrix-site-serve --data DIR --addr ADDR --template T serves HTTP (Axum); --mcp / --acp / --a2a expose the same desks over stdio protocols instead. In production Caddy terminates TLS and reverse-proxies localhost, and systemd keeps both alive. No external database, queue, or cache — state is files under --data.

Runtime topology

Crates

Crate Responsibility
site-serve Composition root: router, shop, atelier, console, CLI, protocol shims
site-edge Holds, releases, the admission gate
site-concierge Grounded answers, voice memos, miss signals
site-commerce Stock, prices, held orders, cashier ledger
site-growth Content desk, lead magnets, schedules, triggers
site-inbox Visitor signals the site couldn't answer
site-defense Request screening, guardrails, freeze switch
site-release Versioned releases, go-live cards
site-migrate Crawl → WorldSpec → parity-scored candidate
site-host Tenancy: one binary, many shops
site-mail Newsletter queue

The agent runtime comes from ADK-Rust: adk-kernel schedules always-on work, adk-harness governs autonomy, adk-runner executes turns under budgets with action ledgers, adk-model serves Meta Muse, adk-session persists sessions, and adk-server runs durable background work.

State: one SQLite file per concern

--data holds cards.db, commerce.db, corpus.db, defense.db, growth.db, kernel.db (durable runs), releases.db, sessions.db, tokens.db, plus the atelier/ portrait directory and migrate-jobs/. Each store opens its own file (mode=rwc) and owns its schema; nothing joins across files.

Models

All agents run Muse through Baseten (BASETEN_API_KEY, BASETEN_MODEL, BASETEN_BASE_URL). --dry-run replaces the model with a scripted stand-in, so the full suite runs offline and free. Portraits and voice memos call Meta's image and transcription APIs directly (META_API_KEY); a deterministic house composer covers the atelier when the model is unreachable.

Governed tool path

Every tool call — concierge, atelier, desks, CLI — dispatches through one authorization path with per-call budgets and a shared action ledger. Anything above the hold line (charge, publish, admit) becomes a held card instead of an action; an owner's decision binds to the call's fingerprint (agent, origin, tool, canonical args), not its ID. The freeze switch halts all dispatch at once.

Autonomy loop

notice → draft → hold → approve → act → learn

Signals (visits, questions, misses, sales) plus a minute worker wake the desks; Content Editor, Lead Gen, Site Agent, and Defender draft; drafts hold as cards; the owner approves in /console; approved cards execute under budget; misses feed the next draft cycle. The desks keep waking on kernel schedule with zero visitors connected.

Machine protocols

Agents never scrape: a public CLI with verified tokens (tokens.db), an MCP server, ACP, and A2A v1.0.0 serve Shop Assistant, Sales Rep, and Feedback Agent. Internal desks (content, reputation, chief of staff, security) are owner-only and unreachable by token.

Auth

Humans: the owner signs /console with a cookie session (OWNER_PASSWORD; the public demo prefills the documented demo password). Machines: per-stall verified tokens with no console access. Defense screens every request before routing.

License

Apache 2.0 — see LICENSE.

About

Agentrix Website — the world's first Website 4.0 for the agentic era. By Zavora Technologies.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages