A Shopify build where agents took the tickets and a review layer decided what shipped.
On a replatform for a direct-to-consumer pet brand, a team of AI agents handled intake, triage, building and reply drafts. I designed the supervision, the independent review, and the limits on what they could touch.
What needed
solving.
Client requests on the replatform needed turning around faster than our previous, slower pipelines allowed. I wanted agents to do the contributing work (intake and triage, building, drafting replies) without handing them anything I couldn’t check or undo.
What I decided,
and why.
- 01
A supervisor. A project-manager agent checked the tickets every few hours and restarted anything that had stalled, usually because an agent ran out of quota or memory, or got stuck in a conflicting action with another agent.
- 02
An independent reviewer. A QA agent checked each change against its spec and sent it back to the building agent when it fell short, so weak first drafts were reworked before they reached me.
- 03
Containment in the infrastructure, not the prompt. Agents worked freely in their own worktrees and a staging copy of the store, but could not write to the live store or to the main, live and staging branches; finished work merged into one second-level branch.
What shipped,
and what isn’t proven.
Agents carried work from intake to a merged, pre-reviewed change, and I reviewed the result each day in a preview theme, looking for what automation can’t see: work outside what was tested, the wrong thing verified, taste and client context. Not yet proven: whether the setup travels to another account.