Vibe coding and AI agents classes in Bridgwater, Somerset

One Overpass request collects every building outline and street that OpenStreetMap holds for a box in Bridgwater measuring 999 m by 1,001 m. The learner lays a grid of 2 m cells over it, 249,500 cells, and marks the 58,365 that fall inside one of the 2,085 building outlines. That is the answer sheet, and the agent never sees it. The agent travels 17.45 km of streets in the box, stops every 5 m, and at each of its 3,149 stops fires 72 beams in a circle, each reaching 30 m. A beam reports the first building cell it meets. Everything the beam crossed on the way counts as a "pass" for those cells, and the cell where it stopped gets a "hit".

With a perfect sensor the limits of the task show at once. The agent observes 124,723 cells, exactly half of the grid, and finds 13,642 wall cells. That is only 23.4% of all building cells, because a sensor on the street sees the faces of buildings and nothing of their insides or backs. The other half of the grid stays unknown, and unknown is not the same as empty. Then the sensor is made realistic. One beam in ten returns a false early echo. One in ten goes through the wall without noticing it. The rest misjudge the distance by about a cell either way.

Trusting the latest reading is the worst plan by every measure. A single stray beam repaints a cell that many earlier beams had agreed on, and a quarter of the real walls are wiped away. Counting votes fixes most of that: phantoms fall from 2,384 to 298. But plain counting treats a hit and a pass as equal, and with a jittery range a wall cell is often passed as well as struck, so a fifth of the walls still vanish. The log-odds score gives evidence different weights. A hit adds 0.85 to the cell's score and a pass takes away 0.4, and the cell is called a wall while its score is above zero. That keeps 92.1% of the walls. The price is 1,420 phantoms. Neither of the last two rows is simply right. The weights are a decision about which mistake costs more, driving into a wall the map forgot or steering round one that was never there, and the agent's designer has to make it.

Bridgwater teenagers who have written that update rule will not build an AI agent that swallows its most recent input whole. It is a lesson learned by coding the agent and watching it fail, which is why learning to code still matters in 2026. The longer argument is in why teenagers should still learn to code in 2026.