Vibe coding and AI agents classes in Stockton-on-Tees

The learner vibe codes a tiny agent and sends it one message per built-up area, largest first: Stockton-on-Tees 84,815, Billingham 33,920, Ingleby Barwick 23,380, Thornaby-on-Tees 23,350, Egglescliffe 10,250, Yarm 9,600, Wynyard 4,040 and Stillington 1,405. Like a chatbot with a very small context window, the agent can only see its last three messages. Asked "Which is the largest town?", it looks at what it can still see and answers Yarm. Asked for the total, it adds up three numbers and says 15,045. Both answers are delivered with complete confidence, because the agent has no idea what it has forgotten.

The fix is a scratchpad memory: after each message, the agent updates a short summary instead of relying on raw history. Keeping the two largest towns, a count and a running sum is enough to answer the first questions correctly. But when the learner later asks for the third largest, the summary cannot help, because nobody told the agent that question was coming. Ingleby Barwick and Thornaby-on-Tees differ by only 30 people, so the answer genuinely needs the stored detail. Keeping the top three solves it. The lesson is that a summary is a decision about which future questions matter.

The running sum hides a final trap. Adding the published figures gives 190,760, not the borough total of 196,595. Smaller places and rural areas are not in the list, and Wynyard's 4,040 is counted whole even though only 2,799 of its residents live inside the borough. So the learner teaches the agent to report the official borough figure when asked for a total, and to describe its own sum as the sum of listed areas. Clear labels matter as much in memory as they do in answers.

Knowing why an AI forgets, and how to design around it, gives Stockton-on-Tees teenagers an edge with every tool they use, and building that knowledge is a strong reason to code in 2026. The longer argument is in why teenagers should still learn to code in 2026.