Preventing Agent Overreach

Overview

AI coding agents have an over-engineering problem. The symptoms are measurable:

The keystrokes got cheaper. The diffs got larger. Agents add features nobody asked for, refactor code that was not in scope, create abstractions for problems that do not exist yet, and "improve" files they were not told to touch. Each of these changes has a cost: review time, regression risk, merge conflicts, and cognitive load for every developer who reads the code next.

The root cause is structural. LLMs are trained on code that is already abstracted, already documented, already refactored. They reproduce those patterns reflexively, regardless of whether the current task calls for them. An agent asked to fix a typo will refactor the function. An agent asked to add a field will redesign the schema. An agent asked to update a dependency will rewrite the module.

This skill enforces a single principle: do exactly what was asked, nothing more.