Research

This paper documents a four-day production run, July 8 to 11, 2026, in which one person shipped sixteen real projects while mostly away from a keyboard. The work spanned client websites, infrastructure operations, a family member's personal assistant, content pipelines, and personal admin. The operating model was not a single powerful agent but a crew: one long-lived orchestrator that plans, dispatches, and verifies but never builds; workers that each own one project; and short-lived subagents routed to the cheapest model that clears the task. Two metered Claude accounts ran in parallel, coordinated through an append-only status log and reached from the road by scannable phone notifications and a tap-to-approve gate for anything irreversible. The human sat above the orchestrator, setting direction and tapping Approve or Deny. This report describes the architecture component by component, presents the campaign as data with real wall-clock samples (including a peak of eighteen concurrent agent sessions), and treats the run's failures as first-class findings. Five of those failures, from a confabulated rogue-orchestrator report to cross-account deliverables that rendered for nobody, drove concrete design changes. No token metering existed per tier during the run, so all cost claims stay qualitative. This is a case study of one operating model, not a benchmark.