130 Designs. Zero Human Pixels. Inside Our Agentic Engineering Demo.
TL;DR
We pointed a fleet of AI agents at a simple challenge: design 130 completely different web experiences, with no human touching a single pixel. Thirteen agents — one orchestrator, ten builders, plus a dedicated critic fleet — took it from a blank folder to a browsable collection you can open right now at pepewebtech.com/demo. This post is the engineering breakdown: the architecture, the rules that made it work, and what it proves about how we build client work.
Why we ran this experiment
Every agency says they use AI. Most mean they have a ChatGPT subscription and a prompt library. We wanted to demonstrate something different: agentic engineering — autonomous agents that plan, produce, critique, and repair work as coordinated teams, with quality control built into the pipeline instead of bolted on at the end.
Design is a good stress test. It can't be faked with a template: someone has to make a hundred coherent decisions about layout, type, color, copy, and interaction, and any weakness is visible instantly. If our agent fleet can ship 130 credible designs with nobody touching the pixels, the same machinery can obviously run your marketing site.
The constraint stack
Each of the 130 designs had to satisfy hard constraints. No exceptions, no partial credit:
- One self-contained HTML file. Zero external assets — no CDN scripts, no web fonts, no stock photos. Every visual is CSS, inline SVG, canvas, or vanilla JavaScript. Each file works offline, forever.
- A named design movement. Not "make it look nice" — Swiss International Style, Art Deco, neo-brutalism, Bauhaus, vaporwave, risograph print, blueprint drafting, and a dozen more. The style had to be recognizable within two seconds of opening the file.
- Genuinely working interactions. Pricing toggles recalculate. Games have win states. Calculators compute correctly. Filters filter. A broken button counts as a failed design.
- Real content. Invented brands, plausible metrics, actual copy. Lorem ipsum anywhere meant automatic rejection.
- Accessible markup. Semantic HTML, labeled controls, keyboard-reachable interactions — because that's our standard on client work too.
The fleet architecture
The pipeline had five stages, and every stage was executed by agents:
| Stage | Who | What happened |
|---|---|---|
| Spec | Orchestrator | Built a deterministic matrix — 13 product categories × 10 subjects × 20 rotating design movements — so all 130 briefs were unique by construction, then froze it as a machine-readable manifest. |
| Design & build | 10 builder agents | Each builder received 10 briefs and worked in parallel, writing every file from scratch against the manifest. |
| Self-critique | Same builders, critic role | After building, each agent audited its own output: requirements met, style fidelity scored, defects logged, fixes applied — with the critique saved as structured evidence next to every file. |
| Adversarial QC | Critic fleet + scripted gate | Fresh agents opened the finished files cold and tried to break them — broken JS, unreachable states, contrast failures, spelling. A scripted gate simultaneously enforced mechanical checks on all 130 files. Nothing shipped until both passed. |
| Ship | Orchestrator | Integrated survivors into this website, wrote this case study, and deployed — still without touching the pixels. |
The rules that made it work
Determinism beats inspiration
The single best decision was generating the assignment matrix mechanically instead of brainstorming topics. Uniqueness was guaranteed by construction, coverage was provable, and any agent could be handed a slice of the manifest and trusted not to collide with another. Creative freedom lived inside tight rails — that's what let thirteen agents work in parallel without producing 130 near-duplicates.
Critique is a separate job from creation
Builders critiqued their own work first, but the decisive pass came from critics who opened files cold with no investment in them. Self-review catches typos; adversarial review catches self-deception. Both layers found real defects the other missed.
Evidence or it didn't happen
Every critique was written to disk as structured JSON: which requirements passed, what defects were found, what was fixed, what risks remain. The orchestrator didn't accept "looks good" as a report. That audit trail is half the point of agentic engineering — quality claims you can verify, not vibes.
What this means for your business
Strip away the spectacle and this demo is a proof of process. The identical pipeline — orchestrated agents that draft, adversarially review each other, and ship with an evidence trail — is how PepeWebTech produces deliverables: landing pages, site sections, campaigns. The practical benefits for a small business are concrete:
- Iteration speed. Ten parallel variations of your page concept instead of one, reviewed and ranked, in the time a traditional workflow spends on the first draft.
- Quality control that actually controls quality. Every deliverable passes an independent critic before you ever see it.
- No blank-page tax. The fleet starts from a frozen spec, not a meeting about what the spec should be.
Judge for yourself
The whole collection is live — filter it by category or style, open any piece, click its buttons. Start with the gallery of all 130 designs, and if you want this machinery pointed at your website, book a call.