This is the first dispatch in a series with one rule: everything reported here happened on our own operation, on our own platform, and is in the ledger. No composites, no hypotheticals, no customer anecdotes borrowed for marketing. When we cite a number, it is queryable.
The operation: a multi-location retail network — stores, a warehouse, online storefronts, point of sale, delivery — that Studio713 owns and runs. It is the reason the lab exists. Before Turbine had a name, it was the machinery we built to stop doing our own operations by hand.
What the agents actually do
Overnight, while no one is watching: inventory reconciles against movement records, and anomalies become flagged proposals rather than silent drift. Support tickets are classified, drafted, and — under threshold — resolved. Campaign drafts are generated from what actually sold, held for human approval. Releases that soaked in staging walk the fleet behind health gates. In the morning, a person reviews decisions, not chores.
The honest part, because that is the point of a dispatch: agents also fail. They have proposed markdowns on products that were selling fine, drafted messages in the wrong register, and once escalated the same incident three times because a heartbeat check was too eager. Every one of those failures is in the ledger too — which is precisely why they were caught, bounded, and fixed instead of compounding in the dark.
The difference between an agent-run operation and an automated mess is not that nothing goes wrong. It is that everything that goes wrong is on the record, bounded to one environment, and reversible.
Why publish this
The loudest voices in this industry are commentators — reporting what the giants announced this week. We hold a different position: we operate. A lab that runs its own company on its own platform does not have to argue that the last mile of AI is solvable. It has the receipts. Future dispatches will publish them: deploy counts, gate outcomes, reversion rates, incident writeups — as the evidence accrues.