Vibe coding and AI agents classes in Cwmbran

The network is invented and has nothing to do with Cwmbran's real roads; it is the textbook example used to explain Braess's result. Four thousand drivers travel from Start to End. The top route runs Start to A, which takes one minute for every hundred cars on it, then A to End, a fixed 45 minutes. The bottom route mirrors it: Start to B, a fixed 45, then B to End at one minute per hundred cars. The learner writes each driver as a small agent with one job: look at yesterday's journey times and, if another route was quicker, consider switching to it.

With only the two routes, and one agent in ten on a slower route switching each round, the crowd settles near 2,000 on each route and every trip takes about 65 minutes. Then the learner adds a free link from A to B that takes no time at all. A driver can now go Start to A, across, then B to End. Nobody is forced to use it.

The first switchers really do gain, and for one round the average trip is slightly shorter. But each car on the shortcut also loads both busy halves of the network, and as more agents follow, the other routes get slower too, which pushes even more agents onto the shortcut. By round 49 almost everyone uses it and the average trip is 79.77 minutes, nearly 15 minutes worse than before the link existed. No agent can do better by switching back alone, so nothing changes from there. Economists call such a state a Wardrop or Nash equilibrium.

Two more runs finish the lesson. If a single coordinator assigned routes, a grid search finds that sending 500 cars over the shortcut and 1,750 along each other route gives 64.69 minutes, better than either result above. The ratio between the selfish outcome and the coordinated one is the price of anarchy. And if every agent reacts to the same information at the same moment, the crowd lurches: with no shortcut and everyone switching at once, all 4,000 cars pile onto one route, then all onto the other, and every trip takes 85 minutes, round after round.

For a Cwmbran teenager, seeing a whole system and not just one clever agent is what lets them direct AI rather than be carried along by it, and the way in is writing the code yourself. The longer argument is in the case for learning to code in 2026.