Vibe coding and AI agents classes in Burnley

The learner builds a small table: the five built-up areas the ONS publishes for Burnley borough, and the census total, 94,646. Then they vibe code a simple agent with two tools, one to read the table and one to write it, and give it the instruction "make the parts add up to the total". The five areas sum to 92,160. The agent's obvious fix is to overwrite the total with 92,160. Another version scales every town up by 2.7 percent so they reach 94,646. Both look tidy, both run without an error, and both damage official figures that were correct all along.

The gap is not an error at all. Built-up areas only cover places that are built up, so the 2,486 people, 2.6 percent of the borough, live outside them, in smaller places and more scattered homes. The agent could not know that from the numbers, and it did not ask. So the learner redesigns its permissions. The original file becomes read-only. The agent may only write to a copy, every change goes into an audit log with the old and new value, and any change to a published figure is blocked outright. Faced with the same instruction, the sandboxed agent proposes adding a clearly labelled derived row and waits for a person to approve it.

The final version also explains itself: it reports that the parts and the total come from different ONS products, that the difference is expected, and that it has changed nothing official. That behaviour, act cautiously, keep originals untouched, log everything and ask when unsure, is exactly what careful teams build into real AI agents that can edit documents, databases and code.

Burnley teenagers who can build AI agents that are safe to trust will be valued wherever AI is used, which is why coding still matters in 2026. The longer argument is in why teenagers should still learn to code in 2026.