Vibe coding and AI agents classes in Crookes, Sheffield

The learner downloads the walkable streets and paths around Crookes from OpenStreetMap, 197.5 km joined at 11,237 points, and a grid of heights from the European EU-DEM elevation model, which runs from about 83 m to 263 m across the area. Every stretch of path gets a slope in each direction. The flat agent assumes 5 km/h everywhere and plans the shortest route. The hill-aware agent uses Tobler's hiking function, in which speed is 6 times e to the power of minus 3.5 times the slope plus 0.05, so a climb and the matching descent cost different times, and plans the quickest route. Both are sent on 300 random trips of at least 600 m.

The surprise is how little the plan changes. The hill-aware agent takes a different route on 37.7% of trips, but those routes are typically only 0.4% quicker, because on a hillside most alternatives climb about the same amount: the median climb is 53 m on the shortest route and 52 m on the quickest. What changes is the prediction. The flat agent's estimate runs a median 9.7% short, and it cannot see that the same walk takes 10.2% longer one way than the other, or 28.8% for the hilliest tenth of trips. Knowing about hills barely improves where the agent goes and greatly improves what it tells you to expect.

Crookes teenagers who have given an agent a better picture of their own hills know what to demand of AI planners, and they learned it by coding one. The longer argument is in why teenagers should still learn to code in 2026.