Vibe coding and AI agents classes in Wythenshawe

The map holds 9,726 building outlines for the area. Volunteers have given a specific type to 3,027 of them (2,858 homes and 169 others) and left 6,699 as plain "building". The learner writes seven labelling functions in Python. Each looks at one thing, the footprint, the corner count, a name, a shared wall, a nearby shop marker, and either votes or stays quiet. Vibe coding suits this well: the learner states the rule in a sentence, an AI writes the function, and the learner checks it against outlines they can see.

No rule covers everything. Across all 9,726 outlines, 1,342 receive no vote and 147 receive votes that disagree. Two ways of settling the votes are compared. A majority vote counts hands. A label model, which never sees a single volunteer tag, works out from the pattern of agreement and silence how far each rule can be trusted, and weighs the votes accordingly.

The result is not the tidy victory the textbooks suggest. The label model finds more of the rare buildings, 144 against 119, and has learned something a show of hands cannot: that a rule staying silent is itself a clue. But it raises more false alarms, and on plain accuracy the majority vote is ahead. Averaging the share of homes and the share of non-homes each method gets right gives 90.8% for the label model and 84.5% for the vote. Which is preferable depends on what a miss costs.

A Wythenshawe teenager who has written labelling functions has seen how AI is fed, which is a sound footing for directing AI agents later. The longer argument is in why teenagers should still learn to code in 2026.