The author evaluates how well AI coding agents perform when instructed to use various testing and verification techniques, testing 26 prompt conditions and 4 skills on a Zstd implementation task in Rust. The study aims to understand whether simple instructions to use techniques like TDD, property-based testing, fuzzing, or formal verification actually improve implementation correctness.
Background
As AI coding agents become more prevalent, understanding how effectively they apply software engineering best practices like testing and verification is critical. This study builds on prior observations that software quality may be declining despite easier access to quality tools.
- Source
- Lobsters
- Published
- Sep 8, 2026 at 12:17 AM
- Score
- 6.0 / 10