EXP 024 TESTING

A verification agent for AI code

RESEARCH APR 2025 Claude Code

Can a dedicated agent catch the silent failures that slip past a quick human review?

Built an agent whose only job was to try to break the generated code and report what failed.

Ran it against real output from another build session.

Caught three silent bugs

AI-generated code still needs verification. An adversarial checker is cheap insurance.

Wire the checker into the build loop so nothing ships unverified.