Vibe coding and AI agents classes in Southall

Street and path data for a rectangle over Southall come from OpenStreetMap, simplified to 3,058 junctions and 218.7 km, of which 832 junctions are dead ends. Python picks 150 random journeys at least 800 m apart; the typical shortest walk is 2,840 m. The agent grows partial routes one junction at a time, scores each by distance walked plus straight-line distance still to go, keeps only the k highest-scoring after every step, and never walks back through a junction it has already used on that route.

A beam of one almost always fails: it follows the most promising street straight into a cul-de-sac with nothing left to fall back on. Widening the beam rescues more journeys, and at 100 it completes 149 of the 150, with routes typically 2.1% longer than the true shortest. It never guarantees the shortest; even at width 100 it found the exact optimum on only 14 journeys, because good routes that looked unpromising early were thrown away. For comparison, A* search, which remembers every junction it has seen, found the shortest route on every journey after examining a median of 438 junctions. On a town map memory is cheap, so A* wins. Beam search earns its place where remembering everything is impossible, such as choosing the next word from tens of thousands, step after step.

A Southall teenager who has watched a greedy agent walk into a cul-de-sac builds agents that keep their options open, and coding is how that lesson sticks. The longer argument is in why teenagers should still learn to code in 2026.