Over the past six months, my SEO content pipeline became fully unattended. Every morning the system picks a topic, runs deep research, writes the article, generates the cover image, publishes it, and then sends me an approval card on Telegram.
The only thing I do is tap ✓ on my phone.
Why "unattended" is the bar
Most people still use AI by typing questions into a chat box. But the real leverage isn't in conversation — it's in pipelines: break a recurring workflow into standard steps, let agents make the judgment calls at each step, and keep humans only at the decision points that matter.
My rules of thumb:
- If I do something three times a week, it deserves a pipeline
- Every step needs a failure fallback — a silent 3 a.m. crash helps no one
- Minimize human checkpoints, but never take them to zero — brand reputation doesn't get delegated to a model
What the system looks like
The whole thing runs on a Mac mini:
- Scheduler: launchd fires the pipeline at a fixed time daily
- Research: automated deep research producing a structured report
- Writing: brand-voice calibration + SEO/GEO dual optimization
- Publishing: straight to the Content API, covers hosted on an R2 CDN
- Notifications: a Telegram bot narrates every step and sends the approval card
The deepest pothole: macOS TCC permissions silently block network requests inside scheduled jobs, with nothing in the logs. The fix is wrapping the job in a node launcher.
The numbers
Two months in: publishing cadence went from one post a week to one post a day, while my own time dropped from six hours a week to twenty minutes.
That's what I mean by: one person plus a fleet of agents is a company.