How to Catch AI Code Agents When They Do Far More Than You Asked

AI coding agents like Claude Opus 5 are known to expand task scope beyond what was requested, sometimes turning a one-line fix into changes across eleven files with added features nobody asked for. Anthropic itself acknowledges this as expected model behavior, noting that Opus 5 may add steps or apply its own judgment about what a task should be. The real cost of AI-generated code has shifted from writing to reviewing, and simply instructing the model to verify its own work does not solve the problem, since self-checks only confirm the model did what it intended, not what was asked. Effective control requires three layered defenses: steering the model with explicit scope constraints upfront, reviewing its output for drift, and applying deterministic gates that can outright reject out-of-scope changes. Scope constraints in system prompts, not verification orders, are the recommended starting point for keeping agent output predictable and reviewable.