Vibe coding and AI agents classes in Sutton Coldfield

First the learner downloads the OpenStreetMap data for a rectangle over Sutton Coldfield in 20 tiles, slowly and politely, and records how big each download was and how long it took. The tiles vary a great deal: the smallest is 0.95 MB, the largest 11.4 MB, and the whole town comes to 112.0 MB and 54.9 seconds of downloading.

Then comes the experiment, which is a simulation. We invent a server rule, and it is ours, not OpenStreetMap's: a bucket holding 20 MB of allowance that refills at 1 MB per second, with any request bigger than the current allowance refused. Seven agents replay the same 20 real tiles against that rule. Because 112.0 MB must pass through a bucket that starts with 20 MB and gains 1 MB a second, no agent can possibly finish in under about 92 seconds.

The impatient agents finish almost as quickly as is possible, but only by being refused again and again: 39 times for the one that retries every second. Real services notice that, and respond with longer blocks or bans. The cautious agent that always waits 10 seconds is never refused, and takes nearly three times as long as it needed to. The agent that models the bucket itself, waiting exactly until enough allowance has dripped in for the next tile, finishes in 93.2 seconds, within about a second of the limit, and is not refused once. Knowing the rule beats both guessing and grovelling.

Building agents that respect limits is how Sutton Coldfield teenagers learn to be trusted with real systems, and that starts with writing the code themselves. The longer argument is in why teenagers should still learn to code in 2026.