An AI content pipeline with one human in the loop
Four AI agents research, generate, schedule, and post content around the clock. Pearl Studio Pro is the one page that shows what the fleet is doing, replaces a text-message approval loop with a gallery picker, and keeps a single human approval in the loop.
- Role
- Stack
- Years
- Status
Four tools, four tabs, and approvals by text message
The pipeline worked, but watching it meant four disconnected surfaces, and approving a batch meant typing replies into a chat. One page was the target.
Everything that varies is data, not code
Agents, pipelines, tasks, actions, and platforms are declared descriptors. The server validates the whole registry at boot and refuses to start on an inconsistent one, cycle checks included.
Watchers for speed, reconciliation for truth
File watchers and server-sent events deliver millisecond updates; a 60-second reconciliation loop re-reads every state file and is the declared authority when they miss.
One gallery, one gesture, one human gate
A picker shows every option per concept and auto-advances through the batch. Each decision is written atomically into the pipeline's own workspace, with the old text path kept as a fallback.
An audit log, four charts, and a retention policy
Every action lands in a filterable audit log with CSV export; metrics cover cost, duration, success rate, and an error taxonomy; six retention policies run nightly.
Built in seven phases by a governed agent fleet
651 commits, 86 logged architecture decisions, and 2,601 automated tests, with more test code than product code and a human sign-off at each phase gate.
The result: 72 hours unattended, one approval per cycle
The pipeline ran three days unattended to pass its gate, and runs since on a 15-minute publishing cadence. The only human step is the approval.
