Category Agent Fleet
One buyer, one merchandiser, one category manager — running a 4,700-product catalog. Instead of hiring, I designed 11 agents to do the work that was repeatable, so the team could focus on the work that wasn't.
Competitive pricing, catalog quality checks, and new product discovery — all running daily and weekly without human input. Built in three weeks to first usable version.
The Problem
Hewyn runs a 4,700-product supplement catalog on a lean team. At that size, three categories of work become impossible to do manually at the right frequency.
Pricing — competitors on iHerb and Amazon.ae update constantly. Without daily checks, you're either leaving margin behind or losing on price without knowing it. Product discovery — reviewing new SKUs every 28 days means acting on stale data; brand catalogs update continuously. Catalog quality — missing images, short descriptions, absent reviews across thousands of products compound invisibly until they show up as lost conversion.
"The question before hiring was whether the work needed more people or a different design. Most of it had rules. Rules can run automatically."
A team of buyers and analysts costs salary. The harder constraint is cognitive: no one holds 4,700 products in context simultaneously. An agent doesn't forget what it checked yesterday.
How It Was Built
Interview the team
Sat with the buyer, merchandiser, and category manager separately. Mapped every task they did regularly — what they checked, how often, what the output was, what happened with it.
Identify what has rules
Filtered for tasks with repeatable logic: pricing comparisons, quality criteria, discovery patterns. These were candidates. Tasks requiring judgment stayed with the team.
Design 11 agents with distinct scopes
Each agent owns one area. Separate scopes prevent overlap; clear outputs make QA tractable. Agents write results to shared Redis keys so any other agent — or a human — can read them.
Build the learning loop
Agents improve from team feedback. Approvals, rejections, and manual overrides feed back in. The system gets better at what the team actually wants, not what was assumed at build time.
The 11 Agents
Before and After
The tasks the agents now cover were previously done manually — price checking, quality reviews, discovery research. The team's time shifted from checking to deciding.
"Agents that learn from feedback behave differently from ones that don't. The first version made mistakes. The current version makes far fewer — because every correction was a training signal."