TableKeeper: A restaurant booking app by AI agents

Tablekeeper is a restaurant reservation service, built entirely by four Claude Code agents working together in one BAND room: a coordinator that plans and accepts, an implementer that builds the back end, a frontend agent that builds every screen, and a reviewer that checks everything and never fixes. I'm Vivek, entering solo from Melbourne, in the Tablekeeper track. The hackathon's hard rule is that a table must never be double-booked, including under concurrent requests, retries and time zones. The factory builds it in four stages: the reservations API, the booking screens with combined tables, booking policies with recurring bookings, and manager table-closure recovery. I sent one message per stage and nothing else. Each seat has one generic mandate file that says how it works and nothing about reservations. The coordinator accepts only a commit the reviewer actually checked, and I re-ran the official harness myself after every stage, outside the factory. All four stages were accepted, with 158 of 158 official checks passing at revision d268821. A fresh clone of the repository passes the same way. The run took about 7 hours 40 minutes from first dispatch to last report. BAND shows $180 for the room, which is BAND's own estimate at list prices. The repository has the four stage folders, the mandates, the unedited room log (room.json), and FACTORY.md, which covers the setup, what went wrong, and measured time and cost. Inputs I supplied before the run: a written design brief, six reference screenshots, five unDraw illustrations and one icon. The agents copied these in unchanged. Limits: the official harness runs only part of the judging tests, so I can vouch for the shipped checks and nothing beyond them. FACTORY.md section 9 lists the rest.