Matteo Vaccari of Thoughtworks argues that when agents write most of the code, humans should specify behaviour through customer-level tests held as tables or text files rather than code. Of the 26 developers he interviewed, most said they did not fully trust their tests, because agents later refactored both the code and the tests. He says data files make unnoticed changes harder, cites early 2026 industry reports of rising defects, and says the approach suits production systems and should be paired with other techniques such as mutation and property-based testing.