`src/content/posts/2026-03-24-notes-from-the-build-7.md`

101 lines

---
title: "How We Fixed 20,000 Product Pages in One Session"
description: "One person and five AI agents audited a 60K+ product catalog, archived 1,099 dead listings, and enriched 20,000 pages — all in a single working session."
date: 2026-03-24
tags: ["ai-automation", "data-quality", "agents", "e-commerce", "catalog-management"]
---

I run a real business and build my own AI tools. No dev team. No six-figure budget. Just Claude Code. Here's today's build.

## The Problem

Product catalogs rot. It happens slowly — a supplier discontinues a line, images expire, specs go stale, metadata drifts. One day you look up and realize thousands of product pages are serving dead links, missing images, or outdated information.

For a catalog with 60,000+ SKUs across dozens of suppliers, manual auditing isn't just tedious — it's impossible. No human team can review 60K pages and make intelligent decisions about each one.

So I built agents to do it.

## Five Agents, One Session

Here's what happened in a single working session:

### 1. The Dead Product Scanner

The first agent crawled the entire catalog and cross-referenced every product against its supplier's live inventory. Products that returned 404s, redirects to category pages, or "discontinued" flags were marked for archival.

**Result: 1,099 dead products identified and archived.** Not deleted — archived. The SEO value and order history are preserved, but customers stop landing on broken pages.

### 2. The Image Coverage Agent

A second agent assessed image coverage across every supplier. One vendor had launched with zero product images — every listing was text-only. The agent pulled images from the supplier's CDN, matched them to the right SKUs, and pushed them to the catalog.

**Result: Image coverage went from 0% to 72% for an entire vendor in under 10 minutes.**

### 3. The Spec Enrichment Assessor

A third agent scanned every product for missing specification data — dimensions, materials, pattern repeats, care instructions. Instead of blindly filling in fields, it categorized products by enrichment readiness: which ones had supplier pages to scrape, which needed manual review, and which were already complete.

**Result: 20,000+ products flagged as ready for automated enrichment**, with a clear priority queue based on traffic and revenue potential.

### 4. The Parallel Research Team

Two more agents ran simultaneously, each handling independent research tasks — one auditing pricing consistency across suppliers, another mapping product taxonomy for better search and filtering.

**Result: All five agents completed their work in parallel**, with no conflicts and no wasted API calls. Total wall-clock time was a fraction of what sequential processing would have taken.

## The Architecture That Makes This Work

This isn't magic. It's a pattern:

1. **Audit first, act second.** Every agent starts by assessing the problem before making changes. You need to know the shape of the mess before you start cleaning.

2. **Never delete, always archive.** Destructive operations are irreversible. Archiving preserves data while removing it from the customer experience.

3. **Parallel when independent.** If two tasks don't share state, run them at the same time. Five agents running in parallel finish in the time of the slowest one, not the sum of all five.

4. **Human-in-the-loop at decision points.** The agents identify and categorize. A human reviews the categories and approves the actions. This is faster AND safer than either pure automation or pure manual review.

5. **PostgreSQL is the source of truth.** Every change gets written to the database first, then pushed to the storefront. If anything goes wrong, you roll back from the database — not from the e-commerce platform's admin panel.

## What This Means for Small Teams

A team of five people couldn't have done this work in a day. They'd need weeks — and they'd make mistakes from fatigue. The agents don't get tired, don't skip rows, and don't accidentally archive the wrong product.

But here's the thing: the agents also don't make judgment calls. They don't know that a product with zero sales but high search volume should be prioritized for enrichment. They don't know that a supplier's "discontinued" page sometimes lies. That's where the human comes in.

The sweet spot is **one person directing multiple agents.** You're not replacing the team — you're replacing the tedium. The human brings context, taste, and business logic. The agents bring speed, consistency, and tirelessness.

## The Numbers

| Metric | Result |
|--------|--------|
| Products audited | 60,000+ |
| Dead products archived | 1,099 |
| Image coverage (one vendor) | 0% to 72% |
| Products ready for enrichment | 20,000+ |
| Parallel agents | 5 |
| Human operators | 1 |

This is what autonomous data quality looks like. Not a dashboard. Not a report. Actual changes, in production, validated and deployed.

## Watch the Video

**[Subscribe to @AgentAbrams on YouTube](https://youtube.com/@AgentAbrams)** for new videos every week.

## Let's Connect

I build AI-powered automation for real businesses — not demos, not prototypes, production systems that run 24/7.

If you're a **founder, entrepreneur, or small business owner** looking to automate operations with AI, let's talk:

- [**@agentabrams on YouTube**](https://youtube.com/@AgentAbrams) — walkthroughs and demos
- [**@agentabrams on X**](https://x.com/agentabrams) — DMs open
- [**@agentabrams on Bluesky**](https://bsky.app/profile/agentabrams.bsky.social) — follow along
- [**goodquestion.ai**](https://goodquestion.ai) — you're here

**Advisory & Board Opportunities:** I'm actively looking to join boards where AI automation can drive real business value. If your company is exploring AI-driven operations, data pipelines, or autonomous agent systems — I'd love to contribute as a board member or advisor. Reach out on any platform above.

---

*Built with [Claude Code](https://claude.ai). Shipped in production. Every day. This is what one founder + AI looks like.*