Agentic workflow · live demo
Point a swarm of AI agents at any website.
Drop in a URL and watch a fleet of specialist agents audit it in parallel, accessibility, content, hierarchy, IA, performance, then adversarially verify every finding, score it against Nielsen's 10 heuristics and the Laws of UX, and synthesize one prioritized report. This is how I design agentic systems.
Every finding is re-opened by a second agent that has to prove it against the page. Anything it can't is dropped.
0 confirmed findings
Top findings
Run an audit to see the report.
Demonstration · findings are illustrative of a real agent audit.
Nielsen's 10 usability heuristics and the Laws of UX, scored against the page: what it already applies, where it falls short, and the fix.
Nielsen's 10 usability heuristics
Loading, save, and progress states are visible.
Labels use the user's words, not system jargon.
Every entered state has a visible exit or undo.
Repeated patterns look and behave the same.
Constraints and confirms stop errors early.
Options and context stay visible, not memorized.
Accelerators speed up frequent tasks.
One clear focus per screen, little clutter.
Errors are plain-language with a way out.
Help and empty states guide the next step.
Laws of UX
Targets are large and well-spaced.
Few, well-grouped choices at each step.
Follows conventions users already know.
Long sets are chunked into groups.
Spacing groups what belongs together.
A credible, intentional first impression.
Each row maps to a research-cited check (NN/g, WCAG, Fitts, Hick, Miller, Gestalt). Verdicts here are illustrative of a real audit.
Why design it this way
The orchestration is the design decision.
Parallelism for coverage
One agent reading a whole site misses things and takes forever. Seven specialists in parallel cover more and finish in a fraction of the time.
Verification for trust
A model that only generates findings will also hallucinate them. A second adversarial pass that must prove each one is what makes the output safe to act on.
Human-in-the-loop for judgment
The agents surface and rank; a person decides what actually ships. AI proposes, the human disposes.
The point
I don't just use AI tools. I design the systems that make a fleet of agents fast, honest, and accountable to a human, and then I ship the result.
How I built this
Behind the build, in full transparency.
Designed the orchestration first
I mapped the work as a pipeline before any pixels: scope the task, fan out to specialist agents in parallel, adversarially verify every finding, then synthesize one report. The shape of that flow is the real design decision.
Built it as a React state machine
A single deterministic clock drives the phases. Each agent's progress and every live counter is computed from elapsed time, so it stays smooth, replayable, and varies its findings from run to run.
Kept it honest and gated
This is a demonstration, so the findings are illustrative rather than a live model's output, and custom URLs sit behind an invite code. The goal is to communicate the pattern clearly, not to fake a product.
The pattern is real
I used this same orchestration to audit and ship this portfolio. A fleet of agents surfaced 84 issues, adversarial verification cut that to 78 real ones, and I fixed them. The demo is the visible half of a workflow I actually run.
The live version is a small serverless function that fetches a page and runs the audit with a real model, gated and cost-capped. Happy to walk through that architecture in person.
Thank you
Thanks for reading.
Want one pointed at your product? Happy to run it with you and walk through what it found.
