Why our companies run on AI agents — and what that actually means
Studio X operates its own ventures on agent pipelines we call XOS. Here is what the agents actually do, what they are not allowed to do, and why we built it this way.
"We run our companies on AI agents" is the kind of sentence that usually means nothing. So here is exactly what it means for us, mechanism by mechanism.
What the agents do
Studio X operates real businesses — a spa booking marketplace, a loyalty platform, a media brand. Inside those businesses, agents hold specific jobs:
- Drafting content. A writer agent turns research and real business data into blog drafts. A line editor — deliberately running on a different model family than the writer — reviews every draft against our voice rules before a human sees it.
- Working the feature backlog. One agent turns user signals and GitHub issues into candidate feature specs. Another dedupes and ranks them. A third implements the top of the queue.
- Keeping data fresh. Scheduled jobs sync business data from Google's APIs nightly and propose enrichment updates for admin review.
- Deploying. Staging deploys run every morning at 07:30 with build verification and a health check that reads the rendered page — because an HTTP 200 with a blank page is still a failure.
What the agents are not allowed to do
This is the part that matters. No agent in our system can publish a public word, merge code, or mutate production data on its own authority.
The content pipeline ends at a named human editor whose approval is what publishes a post. The feature pipeline ends at a test gate — 1,400+ automated checks — and the same review a human contributor would face. The enrichment pipeline proposes; an admin disposes.
We did not add these gates because the agents are bad. We added them because authority should be explicit. When something goes wrong at 2 a.m., "who approved this" needs a one-name answer.
Why bother?
Because it works, and because we sell this. A small studio cannot operate three businesses if people do all the repetitive work. Agents drafting, syncing, and deploying is what makes the portfolio possible — and running the machinery on our own companies first is what lets us tell clients, with a straight face, what it can and cannot do.
Every claim in this post is inspectable in our systems: run logs, append-only iteration receipts, deploy records. That is the standard we hold this blog to. If a future post makes a claim without a receipt, call us on it.