The lede
How well do agents use test and verification techniques?
Dan Luu · made the Sunday
Luu ran 26 testing regimes — TDD, property-based testing, Verus, Lean 4, Kani, plain fuzzing — against the same Zstd implementation task, and the default prompt with no testing guidance beat most of them. Telling an agent to do TDD produced twice as many tests that were smaller and more trivial, and formal-methods runs mostly proved vacuous properties like 0 < c ≤ window while the bitstream-reversal bug sailed through. If you are writing an engineering skill file this quarter, this is the study that says a technique handoff is not the same as a capability.