Vibe coding and AI agents classes in Morningside, Edinburgh

The learner builds a walking network from OpenStreetMap, 130.0 km of streets and paths, and finds the points the map uses to label 14 neighbourhoods in the area, from Firrhill and Greenbank to Merchiston and Greenhill. Dijkstra's algorithm gives the walking distance between every pair. The agent's job: choose one of five venues, the places the map tags as suburbs, for a learner who will walk from one of the 14 neighbourhoods. The agent does not know which. For each venue the program works out the average walk if all 14 starts are equally likely, the longest possible walk, and the largest regret, meaning the most extra distance compared with the venue that would have suited that walker.

An agent told to minimise the average walk chooses the Comiston label point, at 1,387 m. So does a cautious agent using maximin, since Comiston's longest walk, 1,990 m from the Flower Colonies, is the shortest of the five worst cases. But an agent told to minimise regret chooses Morningside: its largest regret is 1,839 m, against 1,967 m for Comiston, whose weak spot is a walker from Merchiston who had a venue on the doorstep. A fourth rule, pure optimism, is useless here: every venue has a shortest possible walk of 0 m. None of the rules is wrong. They answer different questions, and the agent must be told which question it is answering.

A Morningside teenager who can name the rule behind a choice will not mistake an agent's confidence for wisdom, and writing such agents is how that judgement is earned. The longer argument is in why teenagers should still learn to code in 2026.