Vibe coding and AI agents classes in St Andrews

The learner downloads the walkable paths and streets of St Andrews from OpenStreetMap and simplifies them to 3,086 junctions joined by 3,890 segments, 172.4 km in all. Every search starts from the point OpenStreetMap uses to label the town. For each of 60 random destinations the program runs three searches: breadth-first, which counts steps between junctions; depth-first, which dives down one path at a time in a random order; and Dijkstra's algorithm, which always expands the nearest point by actual distance and so finds the true shortest route. The median destination is 1,270 m away, 25 junctions from the start.

All three look at a large share of the town before finding a destination only 25 junctions away, because none of them knows which way the target lies. Breadth-first finds routes with the fewest junctions, but fewest junctions is not the same as fewest metres, so its routes ran about 27% long and it found the true shortest in none of the 60 runs. Depth-first is the one to watch: it explored the most, and the route it brought back was typically eleven times the shortest, over three times too long for 91.7% of destinations, because it follows whatever path it happens to be on. An agent that explores depth-first and reports the first route it finds can be wildly wasteful.

A St Andrews teenager who has coded both searches can tell a thorough agent from a hasty one, and that judgement is built by writing code. The longer argument is in why teenagers should still learn to code in 2026.