Vibe coding and AI agents classes in Llandaff, Cardiff

The learner downloads OpenStreetMap data for a rectangle over Llandaff: 2,569 buildings, whose outlines give 19,452 wall segments, plus the streets and paths. Python marks 2,248 points along those streets and paths, roughly 20 m apart. Two points can see each other if they are no more than 120 m apart and the straight line between them crosses no building wall. That test, repeated for every nearby pair, gives each point its isovist: the set of other street points it can see.

From a typical point an observer sees 27 others, about two thirds of those within range. The spread is wide: one point in ten sees 9 or fewer, one in ten sees 51 or more, and the single most open spot sees 80. Then comes the team question. A greedy planner posts agents one at a time, always choosing the spot that adds the most street points nobody on the team can see yet. A careless planner posts them at random.

Careful placement sees about twice as much as random placement at every team size: ten well-placed observers cover more than twenty careless ones. Even so, twenty agents see under half the streets, because 120 m is a short reach and corners hide a great deal. The model is also generous. It knows nothing of trees, garden walls, slopes or parked vans, so real visibility would be lower. Greedy placement is not guaranteed to be the perfect arrangement either, only a good one that is quick to compute.

A Llandaff teenager who has coded an agent's field of view understands why an AI can be sure and wrong at once, and that understanding comes from building, not from prompting. The longer argument is in why teenagers should still learn to code in 2026.