Same task, same starting query, same timer, two different AI coding agents. This session is a head to head, purely in T-SQL: GitHub Copilot and Claude Code each get an identical, bounded data question against a real small database, run one after the other, live, with the clock visible the whole time. Each round's timer is sized to its task, a short one for a simple write up, a longer one for a genuine diagnosis, but always identical between the two tools.

Each round is short and self contained on purpose, so a slow or wrong attempt is not a disaster, it is the result. Two rounds: write a new report query that genuinely spans several tables from a business question, and fix a query that quietly returns the wrong numbers because of a classic star schema join mistake. After each round we look at what actually got written or fixed, whether it finished inside the timer, and whether the result is something you would trust.

Both rounds run against a small, synthetic Oktoberfest sales database, tents, waiters, sales, food, so the queries are real and the questions being answered are at least entertaining. No cherry picked runs and no do overs, whatever each agent produces before its timer runs out is what the room sees and judges.