Black-Box Testing with AI Agents Checks Software Against Business Rules, Not Code

A developer describes an experiment in which AI models were given only a system's business rules, without access to its code, architecture, or schema. The models wrote tests from those rules and ran them against the live application. The result was a list of every point where the application failed to behave as the business specified. The author argues this approach works because the agents cannot tailor their tests to the implementation. The article outlines a black-box pipeline covering rules as input, test case generation, execution, and a findings report. It also notes that this method is not suited to every situation.