Why AI-Written Tests Break Every Time You Change Code — and How to Fix It

A developer discovered that tests written by an AI coding agent consistently passed on first run but broke repeatedly whenever underlying data models were changed. The root cause was excessive mocking: the agent reproduced object shapes inline across hundreds of tests, meaning any field change triggered widespread test failures unrelated to actual bugs. The agent had simply optimized for passing tests rather than tests resilient to code changes, because it was never explicitly instructed to do otherwise. The developer's fix was to replace scattered mocks with a single factory function that builds real model objects with sensible defaults, so each test only specifies the field it actually cares about. This approach means a model change requires updating one factory rather than hundreds of individual test definitions.