Deleting 1,410 safety checks on purpose to see which ones the tests miss
Published
How does this AI workflow work?
- Replaces: trusting a growing agent-written test suite on faith, and being the copy-paste layer between two coding agents.
- Agents removed 1,410 safety checks from the code one at a time and re-ran the tests: 1,297 were caught, 113 slipped through, 91 of those were real gaps and each now has a test pinning it.
- A supervisor passes work between Claude Code and Codex, which landed 12 packages in 4 days (suite at 6,329 tests).
- The same idea was used on AI-written market reports: 28 planted errors (wrong prices, bad maths, invented figures), all 28 caught by an AI reviewer.
- Honest scoreboard included: consulting calls booked 0 of a goal of 3.
- What to connect: Claude Code and Codex with a supervisor that routes work between them, a test runner the agents can call, and a reviewer agent for generated reports.
- A separate model (ChatGPT here) audits live pipeline runs against the governance rules.
- No cost stated.
- Who it's for: solo developers and small teams shipping agent-written code or reports who want proof that their checks actually catch failures.
From the Morning Tokens issue of Nov 9, 2026