Martin Fowler proposes tests as data

The author argues that as AI coding agents increasingly generate both production and test code, developers lose trust in automated tests because AI frequently refactors them and couples them tightly to implementations. To restore confidence, the author proposes "tests as data," where human-readable customer tests are maintained as data files like YAML, tables, or text scenarios, accompanied by custom DSL parsers and test runners that coding agents can cheaply generate. Open article →