Vibe coding and AI agents classes in Coleraine

The learner downloads the public roads of Coleraine from OpenStreetMap and simplifies them to 1,459 junctions joined by 1,601 links, 119.7 km in total, ignoring one-way rules for this experiment. The destination is the junction nearest the point OpenStreetMap uses to label the town. Each junction becomes an agent that remembers one number, how far it believes it is from the destination. In every round, all agents at once ask each neighbour for that neighbour's number, add the length of the road between them, and keep the smallest total. The destination always says zero; everyone else starts at infinity.

From a cold start the agents agree in 41 rounds, and every final estimate matches Dijkstra's algorithm exactly; the farthest junction is 35 links from the destination, so good news needs only a little longer than that to reach everyone. Then the learner deletes the road OpenStreetMap names the A2 Coleraine Bann Bridge. Traffic can still cross by Sandelford Bridge, but 610 junctions now have a longer true route, typically by 2,229 m. Their agents do not know that. Each still hears old, optimistic numbers from neighbours who were relying on it, so estimates creep upwards a little at a time. The plain method takes 663 rounds and over two million messages to settle. With poison reverse, where an agent tells the neighbour it routes through that its own distance is infinite, it takes 167. Reopening the bridge, good news again, takes just 36.

Designing how agents share news keeps Coleraine teenagers in control of the systems they build, and learning to code is where that begins in 2026. The longer argument is in why teenagers should still learn to code in 2026.