Vibe coding and AI agents classes in Lurgan

The learner downloads OpenStreetMap data and lays a 60 by 60 grid of 20 m squares over central Lurgan, centred on the point the map uses to label the town. A square is blocked if its centre falls inside a mapped building (503 squares) and counts as a main-road square if a trunk, primary, secondary or tertiary road crosses it (346). A toy delivery robot must get from near one corner to near the opposite one. Every move costs 1, and landing on a main-road square costs 10 more, standing in for the risk of being there. The twist: a move goes the chosen way only 80% of the time; otherwise the robot slips sideways.

The naive planner finds the cheapest route as if slips never happened, and after each slip simply replans from wherever it ended up. The Markov decision process instead uses value iteration to work out, for all 3,085 reachable squares, the expected cost to the goal and the action that minimises it, slips included. It needed 229 rounds of updates to settle.

Slips make every journey dearer, but the policy that expects them does noticeably better: its average cost is 7% lower, it strays onto main roads 20% less often, and its bad days are less bad, with the worst tenth of runs costing 194 rather than 214. The difference shows in where it walks: of the naive planner's steps off the main roads, 10.7% were on a square right beside one, against 6.4% for the policy, which gives itself room to slip. The model's own prediction of its average cost, 173.6, matched the simulation closely, a useful check that the maths and the code agree.

Planning for failure is what keeps Lurgan teenagers in charge of the agents they build, and learning to code is how that becomes a habit in 2026. The longer argument is in why teenagers should still learn to code in 2026.