> Preview
> Overview
Agent Commerce is a multi-agent AI system that researches, writes, designs, and reviews digital products end to end — from trend research all the way to a finished, sellable listing. Rather than a plain dashboard, the "Command Centre" visualises the whole pipeline as a pixel-art virtual office: each agent is a little sprite working in its own room (Research Lab, Meeting Room, Content Studio, Print Room, Cyber Cafe), so you can watch Scout, Strategist, Writer, Designer, and the rest actually "at work" in real time, alongside live counters for trends found, products built, and items awaiting review. It runs locally on Apple Silicon using Ollama (Gemma) by default, with an optional Claude API key for higher-quality writing on the content-generation step.
> The Agent Pipeline
Each stage is a distinct agent with one job, passing its output to the next:
SCOUT - researches trends: Google Trends (UK+US), Reddit RSS, Etsy autocomplete FILTER - discards low-quality topics (TV recaps, sports, gossip, news junk) STRATEGIST - designs the product blueprint: sections, worksheet pages, target market WRITER - drafts each section (Claude API if configured, otherwise local Ollama) EDITOR - reviews each section pass/fail, triggers one rewrite loop if needed DESIGNER - builds the final PDF and an Etsy-style cover image QUALITY - reviews content quality and safety (e.g. blocks medical dosing claims) DEPLOY - prepares the Etsy listing SIGNAL - writes matching social copy for Pinterest, Instagram, and TikTok
> Notable Details
- The Command Centre is split into tabs — Command Centre, Trend Intel, Products, Activity
Log, and Settings — so I can jump from watching agents work to reviewing what they
actually produced
- Products are locked to a single market (UK or US) per run, to keep tone and spelling
consistent
- Generated PDFs include real, fillable worksheet pages — tables, weekly grids,
checklists, habit trackers — not just marketing copy
- A safety guard blocks the WRITER agent from making medical or supplement dosing claims,
with the QUALITY agent flagging violations
- Cover art uses warm, Etsy-appropriate palettes chosen per product category, deliberately
not the cyberpunk aesthetic of my other projects
- Everything is served through a FastAPI dashboard, with the active model (e.g.
gemma4:26b) and writer backend shown live in the header
> Why I Built This
I wanted to go beyond a single chatbot and actually build a multi-agent system where each agent has a narrow job and a real handoff to the next — closer to how a small production team would work than how a single prompt would. Digital products were a good testbed because the pipeline has a genuine, verifiable end state: a finished PDF and listing, not just a chat reply.