How well do agents use test/verification techniques?

Researchers tested whether specific instructions to coding agents about testing techniques and libraries improve implementation correctness using Zstd implementation as a benchmark. They evaluated 26 different prompt conditions—including test-driven development, property-based testing, fuzzing, and formal methods tools—across Rust implementations, finding that guidance on testing techniques had varying effectiveness. The study explores whether agents effectively use test/verification methods when prompted, despite evidence suggesting that current defaults in software development may not be yielding quality improvements.