How well do agents use test/verification techniques?

The study evaluates whether prompting coding agents with specific testing techniques or libraries improves Rust implementations of Zstd, testing 26 conditions such as TDD, QuickCheck, property-based testing, fuzzing, SMT solvers, and various skills. Results show that most directives—including “Make no mistakes,” formal-method tools, and the popular ECC test skill—do not outperform a default prompt, and TDD underperforms as predicted.