After publishing my piece on letting AI agents run a retail store, the most common question I received: how does it actually work? This post answers that in full technical detail.

Three agents, three different foundation models selected for task fit rather than convenience, one sequential pipeline orchestrated through 𝑮𝒊𝒕𝑯𝒖𝒃 𝑪𝒐𝒑𝒊𝒍𝒐𝒕 𝑪𝑳𝑰, and a live storefront on RedBubble that sources, designs, and publishes its own products without manual intervention. I cover the agent configurations, the prompt engineering, the Playwright automation, the pipeline wiring, and the honest account of what needed iteration before the output was any good.

I also get into two agents I am planning to add next: one for intellectual property and cultural responsibility screening, and one for automated multi-channel marketing across Instagram, Facebook, and beyond.