How well do agents use test/verification techniques?
Summary
This piece analyzes how different test and verification techniques perform when AI agents implement a Rust-based Zstd project. It tests 26 prompt conditions and multiple testing skills, finding that Default instructions often perform better than when prompting agents to use specific techniques, and that formal methods and many testing libraries yield limited improvements. The article also discusses cost, practical implications for agentic coding, and potential directions for improving agent testing through skills or RL-environment strategies.