Multi-Agent Pipeline · Claude Managed Agents · Shipped
STOREFRONT
TEARDOWN AGENTS
01
A web app that runs four AI agents against any online store. You paste a URL, approve each stage, and about an hour later you have an evidence-backed audit, two rebuilt versions of the storefront, and a report ready to send.
WHAT IT IS
02
A real storefront audit takes a consultant hours. Automated ones are fast but hand back generic scores nobody trusts. This gives you the speed of automation with findings you can check, because every claim points to the page it came from.
WHAT IT SOLVES
03
Sales and agency teams who pitch e-commerce merchants and need something specific and true to say about a store before the first conversation.
WHO IT’S FOR
04
HOW IT WORKS
CRAWL records the store: products, prices, cart, photos. No judging.
01
02
AUDIT finds one to three problems, each with evidence. Sometimes none.
DESIGN rebuilds the storefront two ways using the store's own photos and prices, and re-checks AUDIT's claims against the live site.
03
REPORT packages it all into a report, a slide deck and a summary.
04
A person approves each stage before the next one starts.
05
WHAT IT PRODUCES
06
WHY THE OUTPUT HOLDS UP
Evidence or nothing. Zero findings is a valid result.
The agents check each other. On one run DESIGN caught a wrong AUDIT finding and REPORT dropped it before it reached the final report.
Honest about limits. Every report lists what the agents could not see.
07
HOW I BUILT IT
Four Claude Managed Agents, each with its own prompt and one job, chained by files so each stage reads the last one's output.
A Next.js web app on Vercel that runs the pipeline, shows live progress, holds the approval gates, and keeps an archive of finished runs.
A real bug. The app reported that DESIGN produced nothing when the files existed. It was only reading the first page of the file list. Found and fixed the same day.
About $9 per full run.
Claude Managed Agents (Anthropic API) · Next.js · Vercel · Claude Code · Claude Design
08 - Built With