Multi-Agent Audit
Three parallel AI agents that audit websites, product flows, and user experiences from multiple perspectives at once — with video evidence.
At a glance: Ethereal built a tool that reviews a website from three angles at once — design, marketing, and business — and records video of what it saw as proof for every point it makes. A website review that used to take a designer six hours now finishes in under four minutes.
04The Problem
Website and UX audits are historically slow, manual, and inconsistent. Reviewers click through pages, capture screenshots, and type repetitive reports — bottlenecked and subjective.
A thorough website audit ties up a human designer for six hours before findings land.
The same site reviewed by different people yields different findings — no shared methodology.
Most auditors specialize in one lens; nobody synthesizes UX, marketing, and business at once.
The Solution
We built a map-reduce pipeline with three specialized LLM agents — a UX reviewer, a marketing analyst, and a business analyst — running in parallel. A headless Playwright browser scrolls through the product and captures interaction video; the agents analyze the visual evidence independently, reach consensus on findings, and output a structured report.
Capture
Playwright drives the live site and records video of every interaction and state.
Compress
The DOM, visuals, and interactions are distilled into a structured snapshot the models can reason over.
Fan-out
Three specialized agents — UX, marketing, and business — analyze the same evidence in parallel.
Consensus Engine
The agents reconcile their findings into one report — the map-reduce reduce step, shared by every audit.
Human audit vs. agent turnaround
Large sites are clustered into representative pages and analyzed in parallel chunks, then merged.
Playwright records real interactions — hover states, scroll, motion — not just static screenshots.
One specialist judges layout, hierarchy, and whether the experience is clear and usable.
A second reads messaging, value proposition, and brand for whether the site actually persuades.
A third evaluates conversion paths, trust signals, and whether the funnel can close.
Three perspectives reconcile into one set of ranked, actionable findings — not three siloed reports.
Every finding is grounded in captured video, so recommendations point to what the model actually saw.
The Outcome
Distributed LLM analysis at scale, with objective visual proof. Audits that took a human designer 6 hours complete in under 4 minutes with rich video evidence.
Built with
05 technologiesHave an agentic ai problem like this?
Tell us what you're building — we'll tell you exactly how we'd ship it.
From the archive
Describe a multi-leg options strategy in plain English — the copilot constructs it, validates it against risk rules, and executes across accounts.
An agent that sanitizes enterprise change requests, scores compliance, and plugs straight into the ITSM workflow — with a full audit trail and zero PII leakage.
A self-hosted intraday trading cockpit built around one guarantee: your broker keys never leave your machine.
A white-label, multi-tenant CRM with agentic AI baked into every pipeline — customer visibility, role-based workflows, and AI-powered operations in one place.