Vibe coding and AI agents classes in Bognor Regis

Car availability for every one of Arun's 548 output areas comes from Census 2021 table TS045, fetched through the Nomis API. Across 72,631 households, 12,245 have no car or van, and each area's share becomes one number. The areas are shuffled and split into two halves of 274. Because the split is random, the halves differ only by chance; statisticians call this an A/A test, and a fair method should call them "different" about 5% of the time at the usual p < 0.05 threshold.

The simulated agent receives the areas one pair at a time, one for each half, and runs a Welch t-test comparing the two averages. A patient agent waits until all 274 pairs are in and tests once. An impatient agent tests as it goes and stops as soon as p drops below 0.05, announcing that the groups differ. Python repeats the whole thing on 4,000 different random splits.

Each extra look is another chance for random noise to cross the line, and once it has crossed, the impatient agent stops and never sees it drift back. Checking after every pair turns a 5% false alarm rate into more than 35%. When checking every 10 pairs, half the false alarms had already been declared by the 60th pair. There are honest fixes. One is to fix the sample size in advance and test once. Another, if you must look repeatedly, is a stricter threshold: in this setup, demanding p < 0.0075 at each of the 27 looks brings the false alarm rate back to 5.1%. Proper sequential tests do the same job with formal guarantees.

Setting the rules before the data arrives keeps Bognor Regis teenagers in charge of the agents they build, which is well worth learning to code for in 2026. The longer argument is in why teenagers should still learn to code in 2026.