A UK-based Amazon Merch seller with a large catalogue of designs was spending on ads across all of it, with no clear view of which products were earning that spend back.

The core metric wasn't ACoS (the industry standard). It was Royalty Per Ad Dollar — how much actual royalty income each advertising dollar generated. Against the threshold the seller set, many products were falling short, and nothing was identifying which ones or what to do about them.

The team was spending hours manually checking spreadsheets. They needed a system that read the whole catalogue every day and surfaced the decisions, not more data.

Rather than building a monolithic dashboard, I designed a multi-agent architecture where each agent owns one responsibility — but they all share context through a structured observation log in Supabase.

This means the Ad Waste Detector doesn't just flag underperformers — it writes a structured observation that the Pricing Optimizer can read and act on. Every call either of them makes is an explicit rule with a number behind it, which the seller can open and check. That was his requirement, not mine: he wanted thresholds he could inspect rather than a model he had to take on trust.

Within the first week of deployment, the system surfaced wasted ad spend across underperforming products — spend that had been invisible in manual spreadsheet reviews.

The CEO Brief agent now delivers a weekly executive summary directly to Slack, written in natural language with specific recommendations. The team went from spending hours on data analysis to making decisions in minutes.

Each agent's observations become context for the next one. The Pricing Optimizer reads what the Ad Waste Detector found. The Trend Analyzer informs the CEO Brief. They are not nine separate tools, they are one pipeline with a shared record — and because every decision in it is a written rule, the seller can ask any recommendation why.