Why AI agents over-engineer simple requests and how three prompts can stop it

A developer noticed that an AI coding agent consistently responded to a straightforward automation request — sending outbound emails without a human click — with layered architectural proposals instead of concise solutions. After the first prompt, the agent suggested three automation levels, four channels, and multiple risk tiers, far exceeding the scope of what was asked. Two additional rounds of tighter, more explicit prompting were needed before the agent produced a simple 15-line function with a basic kill-switch. The author attributes this behaviour to training data that rewards consultative, comprehensive responses, which the agent applies even when the task is minimal. The key takeaway is that specifying scope, time constraints, and a maximum number of safeguards upfront can prevent AI agents from defaulting to over-engineered platform-style solutions.