DeepDive - How well do agents use test/verification techniques? (Coding agents are surprisingly bad at testing their own code :-)

Large experiment testing coding agents with 26 different testing/verification instructions plus four agent skills, including: TDD, fuzzing, property-based testing, QuickCheck, Lean 4, SMT solvers, Alloy, Verus, differential testing and others. The experiment used Codex with GPT-5.6 Sol at different reasoning levels.

Model → generates code = Not enough Model → implementation → independent verification → external evaluator → correction = Better approach

Summary1: Giving an agent access to a sophisticated engineering technique ≠ the agent knowing how to use that technique effectively. Summary2: Verification architecture becomes as important as generation quality.

How well do agents use test/verification techniques?

PS: Additional interesting article from Dan is about How does programming language affect token efficiency and correctness