First agents with real customers on the other end

My first agents with real customers on the other end, built on n8n, Supabase and the Claude and OpenAI APIs, for Oyelabs' live client projects.

A form submission triggers a second form by email. The pipeline checks Supabase for a lead we already have, and an AI layer sorts each new one hot, warm or cold. From the lead data it drafts a quotation and routes it to a manager, who approves it before it ever reaches the customer. A Stripe workflow and automatic notifications close the loop.

Before a workflow runs, it enters a set of structured Markdown files and moves through them stage by stage. Each stage states its risks, the resources it needs, and its pass or fail criteria. Nothing executes until the stage it is in says it can.

It gave managers back more than 20 hours a week that had been going into checking work by hand.

The model drafts and a person decides. Most of what I have built since runs on that rule.