The SEO Engine
Curran Dye, marketing lead at Supy, shares his screen and tours the SEO agent that has written about 450 blog posts for Supy: what it researches, how it scores its own drafts, and where a human still presses publish.
About 450 blog posts that used to cost roughly $75 each from outside writers. That is around $30,000 of writing budget, and closer to $40,000 to $50,000 once you price the clicks.
Episode notes
What happens in this episode.
Curran Dye leads marketing at Supy, the AI inventory platform used by multi-site restaurant groups. In the first episode of Agent Cribs he shares his screen and walks through the SEO agent that has written about 450 blog posts for the Supy site. The same work used to cost roughly $75 a post from outside writers, so the agent has replaced around $30,000 of writing budget, and by Curran's own count closer to $40,000 to $50,000 once you price the clicks those posts bring in.
There are no slides. The tour starts in the GitHub repo he treats as the agent's brain, one repo per agent, with the skills, the briefs and a process diagram in the README. From there he opens the two internal MCP servers that feed it: Nexus holds every sales call and customer conversation, so the agent writes about what buyers actually ask, and Cortex reads the Supy codebase, so it never claims a feature the product does not have. Claude routines run the job while his laptop is shut, Supabase holds the drafts, and each post lands in its own Slack channel, scores itself out of 10 against 20 checks, and waits for approve or request edits before it goes live on Webflow.
Ricardo pushes on the awkward parts too. What happens to the agent if Curran gets hit by a bus, whether the repo belongs to him or to Supy, whether a marketer should even want the whole team running it, and what broke on the way here, including the weeks when the laptop had to stay open in his bag for the job to run at all.
On screen
- The GitHub repo he treats as the agent's brain: one repo per agent, with the skills, the briefs and a process diagram in the README.
- Two internal MCP servers. Nexus holds every sales call and customer conversation. Cortex reads the Supy codebase, so the agent never claims a feature the product does not have.
- Claude routines doing the work while his laptop is shut, with Supabase holding the drafts.
- A Slack channel per agent, where each post scores itself out of 10 against 20 checks, then waits for approve or request edits before it goes live on Webflow.
Three things to take home
What to copy, and what not to.
- 01
The agent's brain is a repo, not a prompt. Skills, briefs and the process diagram live in GitHub, so anyone on the team can read how it works and change it.
- 02
It writes from the company's own evidence. Sales calls through Nexus, the codebase through Cortex. That is why the posts sound like Supy and never invent a feature.
- 03
A human still presses publish. Every draft scores itself in Slack and waits for approve or request edits. The approval step is the product, not a bottleneck.
The guest
Curran Dye
marketing lead at Supy
Curran leads marketing at Supy, the AI inventory platform used by multi-site restaurant groups. Before Supy he ran growth at Togather in London and marketing at Atom Learning. He builds his own agents rather than buying them, and this one has a public, dated output: the Supy blog.
Run it with Patricia
The same jobs, from Slack or Teams.
You do not have to build the setup yourself. These are the jobs Patricia already runs that come closest to what the episode shows.
Questions
The short version.
What does the Supy SEO agent actually do?
It researches a topic, drafts a blog post from Supy's own sales calls and codebase, scores the draft out of 10 against 20 checks, and posts it to a Slack channel where Curran approves it or asks for edits. Approved posts go live on Webflow.
How much did the blog posts cost before the agent?
Roughly $75 a post from outside writers. Across about 450 posts that is around $30,000 of writing budget, and by Curran's count closer to $40,000 to $50,000 once you price the clicks those posts bring in.
Does a human still approve each post?
Yes. Every draft waits in Slack for approve or request edits. Nothing publishes on its own, and the edit requests are how the agent's briefs get better over time.
What does the agent run on?
A GitHub repo per agent for the skills and briefs, two internal MCP servers for sales calls and the codebase, Claude routines to run the job on a schedule, Supabase for the drafts, Slack for approval and Webflow for publishing.
What is Curran's advice for a first agent?
Two things. Be inquisitive rather than lazy, and read your own files line by line, because the output is only ever as good as the input.
More from the show
Other doors that are open.
Be on the show
Run something worth showing? We would love to have you tour it.
Apply to be a guestWhat is
Patricia?
Patricia is the AI coworker behind the show. It lives in Slack and Teams: you connect your tools once, then just ask, and Patricia picks up the busywork so you can focus on the parts that need a human. Collaboration, not replacement. That is the whole point of Patricia, and it is why this show is about how real people work alongside their agents.