How Uristocrat.com Runs With AI Agents and No Staff

I run a publication called Uristocrat. It covers sneakers, sports, culture, and technology. It publishes daily editorial content across four verticals, manages its own SEO, generates social assets, runs weekly operations reviews, and syndicates to external platforms.

There is no editorial staff. No freelancers. No content calendar managed by a human. The entire operation runs on a harness of six AI agents I built on Claude, connected to Ghost CMS via MCP servers and deployed on Railway.

This post is about how the system works, what it actually does every day, and what I learned about applied AI in the process.

The stack

There are six agents, each a standalone Claude skill, meaning they can be invoked, tested, and debugged independently, without affecting anything else. They share infrastructure (Ghost CMS via MCP, Railway hosting, Amplitude analytics), but they don't depend on each other; rather, they reinforce each other's work.

Daily Roundup Agent. Every morning, this agent scans sources across all four verticals, identifies the stories worth covering, writes a multi-topic briefing, and publishes it to Ghost. It's the heartbeat of the publication. One post, every day, no human writing involved. The prompt architecture is strict: it can only reference facts from sources it pulled, it writes in Uristocrat's editorial voice, and it flags anything it's uncertain about rather than guessing.

Story Researcher. When something breaks across sneakers, sports, culture, or tech, this agent finds it, suggests angles, and drafts a standalone post. Unlike the roundup, which is scheduled, the researcher runs on demand when I want to go deeper on a topic. It handles the research legwork. I bring the editorial judgment on whether the story is worth telling.

SEO Audit Agent. This one logs in to Google Search Console, pulls every non-indexed URL, diagnoses why each page isn't indexed (crawl issues, canonicalization, thin content, missing metadata), applies fixes through Ghost, and requests a re-crawl. It runs weekly. Before I built it, I had 200+ pages sitting in Google's "discovered but not indexed" queue. That number dropped significantly within the first month.

Instagram Kit. Takes any published Uristocrat article and generates a ready-to-post image and caption. The image uses the article's featured photo with branded overlays. The caption is written specifically for Instagram, not just a rehash of the article's intro. Output goes to a review queue, not straight to posting.

Chief-of-Staff Review. A weekly operations report. It pulls content performance data, subscriber growth, engagement metrics, and surfaces prioritized recommendations for the next week. This is the management layer. Without it, I'd be checking five dashboards manually to understand what's working.

Syndication Agent. Formats and distributes posts to LinkedIn, Apple News, and Yahoo Finance with UTM tracking and Amplitude event logging. Each platform gets content formatted for its specific requirements, not a copy-paste.

What makes this work

A few architectural decisions made this viable rather than a cool demo that falls apart in week two.

MCP as the integration layer. Every agent connects to Ghost CMS via the Model Context Protocol (MCP). This means any agent can read, write, tag, schedule, and publish content through one integration point. I didn't build six custom API wrappers. I built one connection, and six agents use it. When I need to change how publishing works, I change it once.

Non-blocking enrichment. The system never waits for AI to serve content. Data from source APIs returns immediately. AI-generated context (blurbs, summaries, editorial matches) runs in the background and gets cached. The next time the page loads, the enrichment is there. This pattern is critical for anything user-facing. You can't have pages hanging while a language model thinks.

Approval gates by default. Every agent that touches publishing has a dry-run mode. The daily roundup generates, but doesn't publish until I scan it. Syndication requires an explicit --publish flag. The Instagram kit outputs to a review queue. Full autonomy sounds compelling until the system publishes something wrong at 6 am and you wake up to it. The right default is "show me first."

Independent failure. If the SEO agent breaks, the daily roundup still publishes. If syndication goes down, the publication itself is unaffected. No single agent failure takes the whole operation offline. This is the benefit of building standalone skills rather than one monolithic pipeline.

What this actually costs to run

The agents run on Anthropic's API (Claude for generation, Haiku for fast tasks like blurb writing). Ghost CMS is on Railway. Total infrastructure cost is under $200/month. The equivalent human operation, one editor, one SEO specialist, one social media manager, even part-time, would run $8,000-15,000/month minimum.

I'm not saying agents replace those people entirely. The editorial judgment, the taste, the decision about what stories matter, that's still mine. But the execution layer, the daily grind of writing, publishing, optimizing, distributing, and reporting, that's handled.

What I Learned

Too many agents, too early. Six agents sound impressive. In practice, the daily roundup and the story researcher do 90% of the useful work. I would have been better off making two agents great instead of six agents functional.

Agents can write, but they can't decide what matters. The best Uristocrat posts are the ones where I had a take, and the agent helped me execute. The worst ones are where I let the agent decide what was interesting. The human editorial layer is still very much necessary.

I've spent a decade in product and growth roles at startups. Simon Data, Narmi, Finery, Katch, StayTuned. At every one, I watched operational problems that AI could solve today. Content operations were manual and slow. Customer communications were templated and generic. Internal reporting that required a human to pull data from 5 sources and write a weekly summary.

The Uristocrat harness is proof that this can be solved with AI Agents in production.

You can reach me at [email protected] to learn more about this setup.