Vibe coding and AI agents classes in Dungannon

The learner downloads the public roads inside a rectangle over Dungannon from OpenStreetMap: 105.0 km, simplified to 972 junctions joined by 1,043 road segments. For each trip the program picks a random start and destination at least 800 m apart by road, plans the shortest route, and then quietly closes one segment somewhere in the middle of that route. The agent does not know until it arrives at the closed road. Then it can either replan completely from where it stands, or repair the plan by finding the shortest way around to any later point on its old route and carrying on from there. As a yardstick, the program also works out the route the agent would have taken had it known about the closure before setting off.

Discovering a problem late always costs something: even a full replan drives a median 8.3% further than a driver who knew all along, and one trip in ten costs over 45% more. The local repair is slightly worse on distance, 10.4% at the median, but it looks at only 117 junctions to find its way round instead of 403, and in 78.7% of trips it lands within 1% of the full replan. A further 83 closures, from 383 attempted trips, left no route inside the rectangle at all. An agent that keeps searching in that situation wastes effort for ever; it has to recognise the goal is unreachable and report back.

Designing agents that recover sensibly keeps Dungannon teenagers in charge of the AI they build, and learning to code is where that design sense grows in 2026. The longer argument is in why teenagers should still learn to code in 2026.