Vibe coding and AI agents classes in Hamilton

The learner downloads OpenStreetMap data for a rectangle about 1.9 km across over central Hamilton and keeps the drivable streets inside it, treating each as two-way and cutting any that cross the edge. Joining up the points where a street merely bends leaves 453 junctions and dead ends linked by 545 street segments, 44.82 km in total. The agent starts at one junction, must travel every segment at least once, and must finish where it began.

A route that never repeats a street exists only if every junction has an even number of streets, so that each arrival can be matched with a departure. Here 420 junctions are odd, including 148 dead ends, where the only way out is back the way you came. The planner pairs the odd junctions up so that the total distance between partners is as small as possible, a step called minimum-weight matching, and repeats exactly those connecting stretches. The greedy agent does no planning: it takes any street it has not yet driven and, when stuck, heads for the nearest junction that still has one.

Even the perfect plan drives 22.99 km twice, because so many streets end in cul-de-sacs or at the rectangle's edge. The greedy agent drives about 12.75 km more than that on average, and never came within 10 km of the plan in 20 attempts. Its early choices feel efficient, yet they leave scattered unfinished streets that it must later cross town to reach. Planning cost a few lines of code and a matching step; improvising cost nearly 19% more driving.

A Hamilton teenager who demands to see the plan first stays the one in charge, and writing code is where that instinct is trained. The longer argument is in why teenagers should still learn to code in 2026.