Online coding and Python classes in Bury

The census age table on Nomis has 19 rows for each place, a total plus 18 age bands. To compare Bury with England, an agent needs all 38 numbers. The learner's first agent treats the data service like a person answering questions one by one: "how many Bury residents are aged 20 to 24?", then "how many in England?", and so on, each a separate web request. It works, and it gets every number right. It also makes 38 requests that, in our run, added up to 80.0 seconds of waiting and 718 bytes of data.

The batched request asks for both places and every age band at once, and the service returns all 38 rows in a single reply. It carries slightly more data, 932 bytes, because each row includes labels, but it finishes in about one second, eighty times faster. The program checks that every number matches the 38 separate answers exactly. Programmers call the slow pattern the N+1 problem: fetching a list, then fetching each item separately. It is one of the most common reasons real software, and real AI agents, feel sluggish or run up large bills.

Finally the learner gives the agent a tool budget: at most five requests per question. The first agent now fails halfway through, so the learner adds a planning step where the agent writes down what it needs before calling anything and groups those needs into as few requests as possible. With the plan in place, the question about Bury uses one request out of five and leaves room for follow-ups. The timing depends on the network, so the learner runs it several times and reports the range, not a single lucky number.

Bury teenagers who can make AI agents both correct and efficient will be ahead of the crowd, which is why coding is still worth learning in 2026. The longer argument is in why teenagers should still learn to code in 2026.