Vibe coding and AI agents classes in Tynemouth

Census 2021 table TS045 arrives from the Nomis API covering every one of North Tyneside's 729 output areas. Adding them up gives the true answer to the question every agent will be asked: 26,773 of 96,231 households, 27.82%, have no car or van. Each simulated agent answers from a sample of just ten output areas, so any one agent can easily be several points out. A crowd answers with the median of its agents' estimates, a middle value that ignores wild outliers. Every setting is repeated 4,000 times to measure the typical error.

When agents look at their own evidence, the crowd really is wiser: the typical error falls from 4.08 points for one agent to 1.31 for fifteen, and the bad days, the 90th-percentile errors, shrink from 8.37 points to 2.67. When the agents share most of their evidence, the crowd barely improves at all. Fifteen agents that each look at the same eight areas plus two of their own miss by 3.67 points on average, hardly better than one agent alone. They agree with each other because they share the same blind spots, not because they are right.

This is the catch in self-consistency for AI. Asking the same model the same question five times, or giving five agents the same documents, produces answers with shared errors, and their agreement can look far more reassuring than it deserves. Combining answers helps most when the answers come from genuinely different evidence, methods or models.

A Tynemouth teenager who knows when many voices really help will build and use AI agents far more wisely, a great reason to learn to code in 2026. The longer argument is in why teenagers should still learn to code in 2026.