With AI-generated code, tests are becoming the primary artifact developers maintain rather than the code itself. AI agents produce plausible but imperfect code, and tests serve as the feedback harness that guides agents toward correctness — functioning like declarative specifications of desired behavior. AI-generated codebases without tests resemble legacy code in the Feathers sense: inherited black boxes. The recommended approach mirrors legacy code strategies like the strangler pattern, layering in tests from the outside in to gradually enforce correctness.
Sort: