Vibe coding and AI agents classes in Weymouth

The ONS divides the Weymouth built-up area into 192 census output areas and publishes a centre point for each; our sum of the household counts for those areas is 25,576. The learner's agent must draw one closed loop that visits all 192 centres, measured in straight lines. It starts with the areas in the order of their codes, a route nobody planned, which is 135.5 km long. Then it repeats one small step: choose two positions in the loop at random, reverse the stretch between them, and keep the change only if the loop got shorter. Because every step leaves a complete loop, the agent can be stopped at any time and still hand over a valid route.

In the median run, the first 100,000 tries cut the loop from 135.5 km to 51.4 km and the next 900,000 left that figure unchanged. Single runs went on finding small gains a little longer: one stalled at try 93,887, and the last successful step in any of the five came at try 162,402. After that every agent was stuck on a route that no single reversal could shorten. The five runs did not end in the same place, either: their final loops ranged from 50.3 km to 55.9 km, because each had wandered into a different dead end. Running longer would not fix that. Starting again would.

Two comparisons keep the result honest. A different, instant method, always walking to the closest area not yet visited, gives 64.0 km with no waiting; the anytime agent only beats that somewhere between 10,000 and 30,000 tries. And on our laptop 300,000 tries took about two seconds, so here patience is cheap. With an AI agent that pays for every step, the same curve decides real money. These are straight-line distances between area centres, not road journeys, and no route here is proved to be the shortest possible.

Weymouth teenagers who can set a budget for an agent and defend it are doing engineering, not just prompting, and that starts with code they wrote and measured. The longer argument is in why teenagers should still learn to code in 2026.